A task planning evaluation method based on thought chain and strategy optimization

Abstract

The application discloses a task planning evaluation method based on a thinking chain and strategy optimization, and comprises the following steps: S1, constructing a planning original image for task planning evaluation; S2, on the basis of the planning original image, performing thinking chain unfolding on a task planning process of a large model to generate a corresponding chain trace profile; S3, constructing a strategy potential field corresponding to the chain trace profile; and S4, on the basis of the planning original image, the chain trace profile and the strategy potential field, constructing an elastic evaluation ruler to uniformly quantify the task planning capability of the large model. The scheme has the characteristics of improving the scientificity, objectivity and application adaptability of the evaluation of the large model intelligent agent.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large model evaluation technology, and in particular to a task planning evaluation method based on thought chain and strategy optimization. Background Technology

[0002] With the continuous development of artificial intelligence technology, large-scale models have been gradually applied to complex business scenarios such as government services, enterprise operations, industrial processes, customer service, and knowledge management. As the actual needs for digital transformation and intelligent upgrading continue to increase, improving the usability, stability, and controllability of large-scale models in real-world business scenarios has become an important research direction. Especially in complex process tasks, large-scale models often need to complete task analysis, subtask decomposition, tool selection, parameter organization, and execution sequence planning based on user instructions. Therefore, task planning capability has become one of the key capabilities affecting the application effectiveness of large-scale models. Establishing a scientific, unified, and practically applicable task planning evaluation method has significant practical significance and application value.

[0003] Currently, large-scale model capability evaluation mainly focuses on tasks such as question answering, text generation, mathematical reasoning, and code execution. However, effective quantitative evaluation methods are still lacking for tool invocation planning capabilities in complex business processes. In practical applications, large-scale models often face not single-step tasks, but complex goals that require multiple rounds of reasoning and multiple invocations to complete. In this process, the model must not only determine the user's intent but also formulate a reasonable invocation flow and select the optimal solution from multiple possible paths. Existing evaluation methods mostly focus on the final output results, making it difficult to accurately reflect the model's true capabilities in the task planning process. Therefore, how to objectively, meticulously, and interpretably evaluate the tool invocation planning capabilities of large-scale models in complex business processes has become an urgent problem to be solved.

[0004] In existing technologies, evaluation methods for task planning or tool invocation capabilities mainly fall into the following categories: The first category is evaluation methods based on standard answer comparison, which compare the model-generated invocation sequence or final result with a preset answer. The second category is rule-based validation methods, which check whether the tool name, parameter format, and invocation legality meet the requirements. The third category is methods based on manual annotation or scoring, where evaluators subjectively assess the planning process and execution results. The fourth category assumes that each task has only one correct path and judges the correctness of the model output accordingly.

[0005] While the methods described above can perform basic evaluations to some extent, they still have significant shortcomings. First, the method based on comparing standard answers is rigid and difficult to adapt to situations in real-world business scenarios where the same objective corresponds to multiple feasible paths, easily misjudging reasonable but differing planning results as incorrect. Second, the rule-based validation method can only determine whether a call is compliant, but cannot evaluate the rationality of the plan or the efficiency of the step arrangement. Third, while manual evaluation methods can consider more business factors, they are costly, subjective, and difficult to reproduce stably in large-scale scenarios. Finally, the method assuming a single correct path deviates from real-world business conditions, failing to reflect the model's flexible planning capabilities under different constraints. Overall, existing methods generally focus only on results, neglecting the process, and only judge right or wrong without comparing merits, making them insufficient to meet the needs of large-scale model task planning evaluation in complex business processes. Summary of the Invention

[0006] Therefore, a novel task planning evaluation method is needed. This method addresses the difficulty in quantifying and evaluating the tool invocation capabilities of large models in complex business processes, as well as the insufficient adaptability of traditional single-standard answer comparison methods. It introduces the concept of thought chains, transforming the originally difficult-to-observe task planning process into an analyzable multi-step reasoning process. This enables fine-grained evaluation of intent decomposition, step design, and tool selection criteria. Furthermore, by incorporating a strategy optimization mechanism, multiple feasible invocation paths are dynamically compared to evaluate the differences in completion effectiveness, process rationality, and strategic merit among different paths.

[0007] To achieve the above objectives, the inventors provide a task planning and evaluation method based on thought chain and strategy optimization, comprising the following steps:

[0008] S1, Construct the planning archetype for task planning evaluation;

[0009] S2, based on the original planning image, expands the thought chain of the task planning process of the large model to generate the corresponding chain trace profile;

[0010] S3, construct the policy potential field corresponding to the chain trace profile;

[0011] S4, based on the planning preimage, chain trace profile and policy potential field, constructs an elastic evaluation scale to uniformly quantify the planning capability of large model tasks.

[0012] In a preferred embodiment of the present invention, step S1 includes the following steps:

[0013] S101 performs joint parsing of user task text, business context, and rule information, and compresses the target intent, constraints, task objects, and business rules into a unified semantic anchor representation, expressed as:

[0014] ;

[0015] in, Indicates the first A semantic anchor vector, Indicates the first A target intent representation vector. Indicates the first Each constraint condition is represented by a vector. Indicates the first Each task object is represented by a vector. Indicates the first Each business rule is represented by a vector. , , and These represent the mapping matrices corresponding to the target intent, constraints, task objects, and business rules, respectively. Represents a nonlinear transformation function. Indicates the first The significance coefficient of each semantic anchor point This represents the significance adjustment coefficient, where 1 indicates that the coefficient is equal to the average value of the coefficient. All one vectors of the same dimension This indicates element-wise multiplication;

[0016] S102 calculates the adaptation relationship between semantic anchors and the candidate tool set, and forms a stable coupling distribution under the combined effect of constraints and tool capability boundaries. The expression is as follows:

[0017] ;

[0018] in, Indicates the first The semantic anchor point and the first The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool, where B represents the boundary mapping matrix between the semantic space and the tool space. Indicates the first The semantic anchor point and the first Risk distance between candidate tools Indicates the first The semantic anchor point and the first The degree of constraint conflict among candidate tools Indicates the first The semantic anchor point and the first The capability matching gain among the candidate instruments is given by α (risk distance adjustment coefficient), β (constraint conflict adjustment coefficient), γ (capability matching adjustment coefficient), and M (total number of candidate instruments). This represents the transpose of the i-th semantic anchor vector. The vector representing the capability of the k-th candidate tool. This represents the risk distance between the i-th semantic anchor and the k-th candidate tool. This indicates the degree of constraint conflict between the i-th semantic anchor and the k-th candidate tool. This represents the capability matching gain between the i-th semantic anchor and the k-th candidate tool. This represents the operation of the natural exponential function;

[0019] S103 identifies key execution nodes in the task execution chain. It quantifies the importance of computational nodes by introducing completion criteria, expressed as follows:

[0020] ;

[0021] in, Indicates the first The importance score of each candidate execution node, where N represents the total number of semantic anchors. Indicates the first Tool index mapping function corresponding to each candidate execution node Indicates the first Task weights for each semantic anchor point Indicates the causal enhancement coefficient. Indicates the first The causal contribution strength of each candidate execution node in the execution chain Indicates the enhancement coefficient of the completion criterion. Indicates the first Consistency score between each candidate execution node and the completion criterion;

[0022] S104 unifies and aggregates the target intent, constraints, tool boundaries, key execution nodes, and completion criteria to form a planning preimage for subsequent evaluation, expressed as:

[0023] ;

[0024] in, Indicates the original image of the plan. This represents the aggregated target intent preimage vector. Let X represent the preimage vector of the aggregated constraint conditions, where X represents the vector formed by the total coupling strength. The tool boundary coupling matrix is ​​formed. Represents the set of critical execution nodes. This represents the preimage vector of the completion criterion. Indicates the first The weight of each semantic anchor in the aggregation of target intent Indicates the first The weights of each semantic anchor point in the constraint aggregation, where N represents the total number of semantic anchor points, M represents the total number of candidate tools, and P represents the total number of candidate execution nodes. Indicates the first The structural representation vector of each candidate execution node Represents the normalization function. This represents the importance score of all nodes. The constructed score vector, This indicates that the highest score is selected. One candidate execution node.

[0025] In a preferred embodiment of the present invention, in step S1, the planning preimage is used to extract the target intent, constraints, tool boundaries, key execution nodes and completion criteria in the user task in a structured manner, and form an evaluation benchmark consistent with the actual business needs.

[0026] In a preferred embodiment of the present invention, step S2 includes the following steps:

[0027] S201, regarding the original plan The task planning step sequence output by the large model Using the preimage projection expansion operator, we obtain the preimage-aligned thought step representation, expressed as:

[0028] ;

[0029] in, Let represent the representation vector of the u-th original task planning step, and L represent the total number of task planning steps. Let the vector represent the thought step of the u-th preimage alignment. This represents the original planning step mapping matrix. The target intent preimage mapping matrix represents the matrix. This represents the preimage mapping matrix of the constraint conditions. Represents the tool boundary mapping matrix. This represents the mapping matrix of key execution nodes. This represents the preimage mapping matrix for the completion criterion. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. Indicates the first The semantic anchor point and the first The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool. Indicates the first The importance score of each candidate execution node Indicates the first The structural representation vector of each candidate execution node Indicates the first Does each candidate execution node belong to the set of critical execution nodes? The selection flag, when the node belongs to hour ,otherwise K represents the number of critical execution nodes. This represents the preimage vector of the completion criterion. This represents the explicit gain coefficient for the u-th task planning step. Represents the normalization function;

[0030] S202, Preimage Alignment Thinking Step Representation Using a multi-track splitting mapping operator, we obtain track-by-track representations of the target recognition trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory, and step connection trajectory. The expression is:

[0031] ;

[0032] in, Let represent the trajectory sub-vector of the u-th task planning step on the z-th type trajectory, where z is the trajectory type number, z=1 represents the target recognition trajectory, z=2 represents the subtask decomposition trajectory, z=3 represents the tool selection trajectory, z=4 represents the parameter organization trajectory, and z=5 represents the step connection trajectory. This represents the step-state mapping matrix corresponding to the z-th type of trajectory. This represents the semantic anchor mapping matrix corresponding to the z-th type of trajectory. This represents the tool capability mapping matrix corresponding to the z-th type of trajectory. In the planning step of the u-th task, the first step is... The contribution weights of each semantic anchor point to the z-th type trajectory Indicates the first A semantic anchor vector, In the planning step of the u-th task, the first step is... The contribution weights of each candidate tool to the z-th type trajectory;

[0033] S203, representing five types of track separation Using the sheet weaving operator, we obtain chain trace sheets that characterize the internal structure of the current planning step and the connection between the preceding and following steps. The expression is:

[0034] ;

[0035] in, Indicates the first The chain trace layer slice vector corresponding to each task planning step This represents the intra-layer contribution coefficient of the z-th type trajectory in the u-th task planning step. In the planning step of the u-th task, the first step is... Class trajectory and the first The weaving coefficient between class trajectories, Indicates the selection marker for the trajectory pair, when hour ,when hour , This indicates element-wise multiplication. This represents the continuation coefficient of the u-th task planning step from the previous task planning step. The vector represents the step-connection trajectory of the previous task planning step. Number 5 corresponds to the step-connection trajectory. hour, Take the zero vector;

[0036] S204, for all chain trace layers And the corresponding five types of trajectory sequences, using the profile aggregation operator, yield the chain track profile Ω, expressed as:

[0037] ;

[0038] Where Ω represents the chain track profile. Represents the target recognition trajectory sequence. This represents the sequence of subtask decomposition trajectories. The tool indicates that a trajectory sequence has been selected. This indicates the parameter organization trajectory sequence. This indicates the sequence of steps connecting the trajectory. Indicates that it consists of all chain trace layers. The constructed layer sequence, The aggregate signature vector representing the chain trace profile. This represents the weight of the u-th task planning step in the profile aggregation.

[0039] As a preferred embodiment of the present invention, in step S2, the chain trace profile is used to characterize the target identification trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory and step connection trajectory formed by the model in the task solving process, and to present the key cognitive links in the planning process in a hierarchical manner.

[0040] In a preferred embodiment of the present invention, step S3 includes the following steps:

[0041] S301, comparing the chain track profile Ω with the planned preimage. Using the path embryo kernel extraction operator, feasible planning path embryo kernel representations around the same task objective are obtained. The expression is:

[0042] ;

[0043] in, Let r be the kernel representation vector of the r-th feasible planning path, where r = 1, 2, ..., R; and R represent the total number of feasible planning paths. Represents the chain trace layer-slice mapping matrix. The step-by-step trajectory mapping matrix represents the connection between the steps. This represents the mapping matrix of key execution nodes. Let represent the assignment coefficient of the u-th task planning step to the r-th feasible planning path. This represents the chain trace layer slice vector corresponding to the u-th task planning step. Let represent the step connection trajectory vector for the u-th task planning step, where number 5 corresponds to the step connection trajectory. Indicates the first Does each candidate execution node belong to the set of critical execution nodes? The selection marker, Indicates the first The importance score of each candidate execution node Indicates the first The structural representation vector of each candidate execution node Represents the normalization function;

[0044] S302, for feasible planning path kernels Chain trace profile aggregate signature and original plan Using the potential ripple projection operator, the attribute vector of the r-th feasible planning path is obtained. The expression is:

[0045] ;

[0046] in, Let represent the attribute vector of the r-th feasible planning path. Let represent the execution efficiency score of the r-th feasible planning path. This represents the step redundancy score for the r-th feasible planning path. This represents the resource cost score of the r-th feasible planning path. This represents the risk level score of the r-th feasible planning path. This represents the constraint satisfaction score for the r-th feasible planning path. Represents the path embryo kernel mapping matrix. This represents the aggregate signature mapping matrix of the chain trace profile. The target intent preimage mapping matrix represents the matrix. This represents the preimage mapping matrix of the constraint conditions. This represents the preimage mapping matrix for the completion criterion. The aggregate signature vector representing the chain trace profile. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. This represents the preimage vector of the completion criterion;

[0047] S303, for all feasible planning path kernels and path attribute vector Using the potential field texture operator, the policy potential field coupling matrix is ​​obtained. The expression is:

[0048] ;

[0049] in, Let represent the potential coupling strength between the r-th feasible planning path and the s-th feasible planning path. Represents the policy potential coupling matrix. Represents the path-nucleus coupling matrix. Represents the path attribute coupling matrix. and Let represent the kernel representation vectors of the r-th and s-th feasible planning paths, respectively. and Let r and s represent the attribute vectors of the r-th and s-th feasible planning paths, respectively.

[0050] S304, for path attribute vectors Coupled matrix with strategy potential field Using the path potential energy summation operator, the path potential energy of the r-th feasible planning path is obtained. and path hierarchy tags The expression is:

[0051] ;

[0052] in, Let r represent the path potential energy of the r-th feasible planning path. The hierarchical label represents the r-th feasible planning path. Indicates the execution efficiency weight. Indicates the redundancy weight of the steps. Indicates the resource cost weight. Indicates the risk level weight. The weights represent the constraints that are satisfied. Indicates the potential coupling weight. Let represent the execution efficiency score of the r-th feasible planning path. This represents the step redundancy score for the r-th feasible planning path. This represents the resource cost score of the r-th feasible planning path. This represents the risk level score of the r-th feasible planning path. This represents the constraint satisfaction score for the r-th feasible planning path. Let represent the potential coupling strength between the r-th feasible planning path and the s-th feasible planning path. and The threshold for hierarchical discrimination is represented, and it satisfies the following conditions: , The hierarchical discriminant function is defined when... Output the optimal path label when When the output accepts path labels, Inefficient path labels are output at times.

[0053] In a preferred embodiment of the present invention, in step S3, the strategy potential field is used to characterize the distribution of feasible planning paths formed around the same planning preimage, and to conduct a unified analysis on the different paths in terms of execution efficiency, step redundancy, resource cost, risk level and constraint satisfaction. Based on this, a path potential energy calculation mechanism is established to measure the comprehensive efficiency level of each planning path in the process of achieving the goal.

[0054] In a preferred embodiment of the present invention, step S4 includes the following steps:

[0055] S401, regarding the original image of the plan Ω-strategy potential field of chain trace profile Using the facet convergence operator, the five-dimensional evaluation facet vector corresponding to the elasticity evaluation scale is obtained. The expression is:

[0056] ;

[0057] in, Represents the five-dimensional evaluation facet vector. Indicates the degree of matching between task objectives. Indicates the completeness of the thought process chain. Indicates the appropriateness of tool usage. Indicates the path potential energy level. Indicates the status of business constraint satisfaction. This represents the preimage mapping matrix. Represents the chain track profile mapping matrix. The tool coupling mapping matrix is ​​represented. Represents the strategy potential mapping matrix. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. This represents the preimage vector of the completion criterion. The aggregate signature vector representing the chain trace profile. Indicates the relationship between the i-th semantic anchor and the th... The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool. Indicates the first The path potential energy of a feasible planning path, The attribute vector representing the r-th feasible planning path;

[0058] S402, evaluation of five-dimensional faceted vectors Strategic potential field Using the task structure parameters and the elastic weight operator, a dynamic weight vector is obtained. The expression is:

[0059] ;

[0060] in, Represents a dynamic weight vector. , , , and The weights correspond to five evaluation dimensions: the degree of matching between task objectives, the completeness of the thought process, the rationality of tool usage, the level of path potential, and the satisfaction of business constraints. Represents the basic weight vector. This represents the task complexity adjustment factor. Indicates the path diversity adjustment coefficient. This represents the direction vector for complexity weighting. Represents the direction vector of diversity weighting. Indicates the first The potential coupling strength between the s-th feasible planning path and the s-th feasible planning path. This represents the normalized weight function;

[0061] S403, evaluation of five-dimensional faceted vectors and dynamic weight vector The comprehensive score is obtained by using the elasticity evaluation scale calculation operator. And the diagnostic vector z, thus forming the elasticity evaluation scale result. The expression is:

[0062] ;

[0063] in, This indicates the results of the elasticity evaluation scale. This indicates the overall score. Represents the diagnostic vector. , , , and These correspond to the weighted deviation values ​​of the five evaluation facets. This represents the baseline threshold vector for the five evaluation facets. This indicates element-wise multiplication. Represents a dynamic weight vector. This represents the five-dimensional evaluation facet vector.

[0064] In a preferred embodiment of the present invention, in step S4, the elastic evaluation scale is used to comprehensively consider the degree of matching of task objectives, the completeness of the thought chain, the rationality of tool calls, the level of path potential energy, and the satisfaction of business constraints, and dynamically adjusts the evaluation weights in combination with the complexity of the task and the diversity of the planned path to form a comprehensive score result.

[0065] Unlike existing technologies, the above technical solution achieves the following beneficial effects:

[0066] (1) This method proposes a task planning evaluation method for tool calling scenarios, which is used to systematically analyze and quantitatively evaluate the planning ability of the model in actual business tasks;

[0067] (2) This method establishes an evaluation mechanism that balances uniformity and flexibility, so that the evaluation results are comparable and can reflect the actual situation in complex business processes.

[0068] (3) This method introduces a thought chain analysis mechanism, which transforms the task planning process from an unobservable to an observable and analyzable reasoning chain, thereby improving the interpretability and refinement of the evaluation process;

[0069] (4) This method combines the idea of ​​strategy optimization to dynamically compare and comprehensively evaluate multiple feasible calling paths, providing a reliable basis for the capability diagnosis, effect improvement and actual deployment of large model agents;

[0070] (5) While maintaining a unified evaluation standard, this method is better adapted to the multiple solutions and process flexibility in real business scenarios, which is necessary to improve the scientificity, objectivity and application adaptability of large model intelligent agent evaluation. Attached Figure Description

[0071] Figure 1 The flowchart is for a specific implementation method. Detailed Implementation

[0072] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0073] like Figure 1As shown, this embodiment provides a task planning evaluation method based on thought chain and strategy optimization to improve the evaluation of tool call planning capabilities in complex business processes. Specifically, it includes the following steps:

[0074] S1. First, a planning preimage for task planning and evaluation is constructed. The planning preimage is used to structurally extract the user's task objectives, constraints, tool boundaries, key execution nodes, and completion criteria, forming an evaluation benchmark consistent with real business needs. This step does not rely on a single answer as a prerequisite for evaluation; instead, it uses the task's planning space and completion requirements as a unified reference, enabling subsequent evaluations to balance standard consistency with multiple business solutions, thereby improving the authenticity and adaptability of task planning and evaluation from the outset.

[0075] S2, based on the original planning model, expands the thought process of the large-scale task planning process into a chain, generating corresponding chain trace profiles. These profiles characterize the target identification trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory, and step connection trajectory formed by the model during task solving, presenting key cognitive links in the planning process in a hierarchical manner. This step transforms the internal planning process, invisible in traditional evaluation, into an analyzable, comparable, and locatable evaluation object, enabling fine-grained evaluation of the accuracy of task understanding, the completeness of step organization, the rationality of tool selection, and the coherence of calling logic.

[0076] S3. Construct the strategy potential field corresponding to the chain path profile. The strategy potential field is used to characterize the distribution of feasible planning paths formed around the same planning preimage, and to conduct a unified analysis of different paths in terms of execution efficiency, step redundancy, resource cost, risk level, and constraint satisfaction. Based on this, a path potential energy calculation mechanism is established to measure the comprehensive efficiency level of each planning path in the process of achieving the goal. Through this step, task planning evaluation is no longer limited to the single path hit judgment, but can perform hierarchical comparison of multiple feasible paths, thereby identifying the optimal path, acceptable path, and inefficient path, and improving the ability of the evaluation results to reflect the differences in real business strategies.

[0077] S4, based on the preimage, chain path profile, and policy potential field, constructs a flexible evaluation scale to uniformly quantify the task planning capabilities of large models. The flexible evaluation scale comprehensively considers the degree of task objective matching, the completeness of the thought chain, the rationality of tool calls, the level of path potential energy, and the satisfaction of business constraints. It also dynamically adjusts the evaluation weights in combination with task complexity and the diversity of planned paths, thereby forming a comprehensive scoring result that is more in line with actual application scenarios.

[0078] In the specific implementation of the above embodiments, step S1 includes the following steps:

[0079] S101 performs joint parsing of the user task text, business context, and rule information, compressing the target intent, constraints, task objects, and business rules into a unified semantic anchor representation. After this processing, the original task description can be transformed from natural language form into structured input that can be directly used in subsequent evaluations, expressed as:

[0080] ;

[0081] in, Indicates the first A semantic anchor vector, Indicates the first A target intent representation vector. Indicates the first Each constraint condition is represented by a vector. Indicates the first Each task object is represented by a vector. Indicates the first Each business rule is represented by a vector. , , and These represent the mapping matrices corresponding to the target intent, constraints, task objects, and business rules, respectively. Represents a nonlinear transformation function. Indicates the first The significance coefficient of each semantic anchor point This represents the significance adjustment coefficient, where 1 indicates that the coefficient is equal to the average value of the coefficient. All one vectors of the same dimension This indicates element-wise multiplication.

[0082] S102 calculates the fit between semantic anchors and the candidate tool set, forming a stable coupling distribution under the combined effect of constraints and tool capability boundaries. After this processing, it is possible to determine which tools are more suitable for each semantic anchor, and to identify capability mismatches and constraint conflicts in advance. The expression is:

[0083] ;

[0084] in, Indicates the first The semantic anchor point and the first The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool, where B represents the boundary mapping matrix between the semantic space and the tool space. Indicates the first The semantic anchor point and the first Risk distance between candidate tools Indicates the first The semantic anchor point and the first The degree of constraint conflict among candidate tools Indicates the first The semantic anchor point and the first The capability matching gain among the candidate instruments is given by α (risk distance adjustment coefficient), β (constraint conflict adjustment coefficient), γ (capability matching adjustment coefficient), and M (total number of candidate instruments). This represents the transpose of the i-th semantic anchor vector. The vector representing the capability of the k-th candidate tool. This represents the risk distance between the i-th semantic anchor and the k-th candidate tool. This indicates the degree of constraint conflict between the i-th semantic anchor and the k-th candidate tool. This represents the capability matching gain between the i-th semantic anchor and the k-th candidate tool. This represents the operation of the natural exponential function.

[0085] S103 identifies key execution nodes in the task execution chain and quantifies their importance by introducing completion criteria. After this processing, the system not only identifies the key execution nodes that determine the quality of task completion but also transforms the completion criteria from abstract descriptions into quantifiable node evaluation signals, expressed as:

[0086] ;

[0087] in, Indicates the first The importance score of each candidate execution node, where N represents the total number of semantic anchors. Indicates the first Tool index mapping function corresponding to each candidate execution node Indicates the first Task weights for each semantic anchor point Indicates the causal enhancement coefficient. Indicates the first The causal contribution strength of each candidate execution node in the execution chain Indicates the enhancement coefficient of the completion criterion. Indicates the first Consistency score between each candidate execution node and the completion criterion.

[0088] S104 unifies and aggregates the target intent, constraints, tool boundaries, key execution nodes, and completion criteria to form a planning preimage that can be used for subsequent evaluation. After this processing, the task no longer relies on a single standard answer for description, but is given an evaluation reference in the form of a structured preimage, thus simultaneously retaining the task completion requirements and multi-path planning space. The expression is:

[0089] ;

[0090] in, Indicates the original image of the plan. This represents the aggregated target intent preimage vector. Let X represent the preimage vector of the aggregated constraint conditions, where X represents the vector formed by the total coupling strength. The tool boundary coupling matrix is ​​formed. Represents the set of critical execution nodes. This represents the preimage vector of the completion criterion. Indicates the first The weight of each semantic anchor in the aggregation of target intent Indicates the first The weights of each semantic anchor point in the constraint aggregation, where N represents the total number of semantic anchor points, M represents the total number of candidate tools, and P represents the total number of candidate execution nodes. Indicates the first The structural representation vector of each candidate execution node Represents the normalization function. This represents the importance score of all nodes. The constructed score vector, This indicates that the highest score is selected. One candidate execution node.

[0091] In the specific implementation of the above embodiments, step S2 includes the following steps:

[0092] S201, regarding the original plan The task planning step sequence output by the large model Using the preimage projection expansion operator, we obtain the thought step representation with preimage alignment. After this processing, each reasoning step in the original planning process can establish an explicit correspondence with the target intent, constraints, tool boundaries, key execution nodes, and completion criteria, expressed as:

[0093] ;

[0094] in, Let represent the representation vector of the u-th original task planning step, and L represent the total number of task planning steps. Let the vector represent the thought step of the u-th preimage alignment. This represents the original planning step mapping matrix. The target intent preimage mapping matrix represents the matrix. This represents the preimage mapping matrix of the constraint conditions. Represents the tool boundary mapping matrix. This represents the mapping matrix of key execution nodes. This represents the preimage mapping matrix for the completion criterion. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. Indicates the first The semantic anchor point and the first The coupling strength between candidate tools Indicates the first A capability representation vector for each candidate tool, where N represents the total number of semantic anchors, M represents the total number of candidate tools, and P represents the total number of candidate execution nodes. Indicates the first The importance score of each candidate execution node Indicates the first The structural representation vector of each candidate execution node Indicates the first Does each candidate execution node belong to the set of critical execution nodes? The selection flag, when the node belongs to hour ,otherwise K represents the number of critical execution nodes. This represents the preimage vector of the completion criterion. This represents the explicit gain coefficient for the u-th task planning step. This represents the normalization function.

[0095] S202, Preimage Alignment Thinking Step Representation Using a multi-track splitting mapping operator, we obtain track-by-track representations of the target recognition trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory, and step connection trajectory. After this processing, the different cognitive stages that were originally mixed in the same reasoning process are deconstructed into five independently analyzable trajectories, expressed as:

[0096] ;

[0097] in, Let represent the trajectory sub-vector of the u-th task planning step on the z-th type trajectory, where z is the trajectory type number, z=1 represents the target recognition trajectory, z=2 represents the subtask decomposition trajectory, z=3 represents the tool selection trajectory, z=4 represents the parameter organization trajectory, and z=5 represents the step connection trajectory. This represents the step-state mapping matrix corresponding to the z-th type of trajectory. This represents the semantic anchor mapping matrix corresponding to the z-th type of trajectory. This represents the tool capability mapping matrix corresponding to the z-th type of trajectory. In the planning step of the u-th task, the first step is... The contribution weights of each semantic anchor point to the z-th type trajectory Indicates the first A semantic anchor vector, In the planning step of the u-th task, the first step is... The contribution weights of each candidate tool to the z-th type trajectory.

[0098] S203, representing five categories of track division Using the sheet weaving operator, we obtain chain trace sheets that characterize the internal structure of the current planning step and the connection between the preceding and following steps. After this processing, the cooperative relationships between different trajectories and the continuity relationships between adjacent steps are jointly encoded into the same layer, enabling the chain track profile to exhibit a layered structural feature, expressed as:

[0099] ;

[0100] in, Indicates the first The chain trace layer slice vector corresponding to each task planning step This represents the intra-layer contribution coefficient of the z-th type trajectory in the u-th task planning step. In the planning step of the u-th task, the first step is... Class trajectory and the first The weaving coefficient between class trajectories, Indicates the selection marker for the trajectory pair, when hour ,when hour , This indicates element-wise multiplication. This represents the continuation coefficient of the u-th task planning step from the previous task planning step. This represents the step-connection trajectory vector of the previous task planning step, where number 5 corresponds to the step-connection trajectory. hour, Take the zero vector.

[0101] S204, for all chain trace layers The corresponding five types of trajectory sequences are used to obtain the chain track profile Ω using the profile aggregation operator. After this processing, the target recognition trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory, and step connection trajectory formed during the large model task planning process can be output with a unified structure, thus forming an analyzable, comparable, and localizable evaluation object. The expression is:

[0102] ;

[0103] Where Ω represents the chain track profile, Represents the target recognition trajectory sequence. This represents the sequence of subtask decomposition trajectories. The tool indicates that a trajectory sequence has been selected. This indicates the parameter organization trajectory sequence. This indicates the sequence of steps connecting the trajectory. Indicates that it consists of all chain trace layers. The constructed layer sequence, The aggregate signature vector representing the chain trace profile. This represents the weight of the u-th task planning step in the profile aggregation.

[0104] In the specific implementation of the above embodiments, step S3 includes the following steps:

[0105] S301, comparing the chain track profile Ω with the planned preimage. Using the path embryo kernel extraction operator, feasible planning path embryo kernel representations around the same task objective are obtained. After this processing, the local inference steps in the chain path profile no longer exist in isolated form, but are reorganized into a path-level representation that reflects the complete planning direction, expressed as:

[0106] ;

[0107] in, Let r be the kernel representation vector of the r-th feasible planning path, where r = 1, 2, ..., R, and R represents the total number of feasible planning paths. Represents the chain trace layer-slice mapping matrix. The step-by-step trajectory mapping matrix represents the connection between the steps. This represents the mapping matrix of key execution nodes. Let represent the assignment coefficient of the u-th task planning step to the r-th feasible planning path. This represents the chain trace layer slice vector corresponding to the u-th task planning step. Let represent the step connection trajectory vector for the u-th task planning step, where number 5 corresponds to the step connection trajectory. Indicates the first Does each candidate execution node belong to the set of critical execution nodes? The selection marker, Indicates the first The importance score of each candidate execution node Indicates the first A structural representation vector of candidate execution nodes, where L represents the total number of task planning steps, P represents the total number of candidate execution nodes, and K represents the number of critical execution nodes. This represents the normalization function.

[0108] S302, for feasible planning path kernels Chain trace profile aggregate signature and original plan Using the potential ripple projection operator, the attribute vector of the r-th feasible planning path is obtained. After this processing, the differences in execution efficiency, step redundancy, resource cost, risk level, and constraint satisfaction among different paths can be uniformly projected into the same attribute space, expressed as:

[0109] ;

[0110] in, Let represent the attribute vector of the r-th feasible planning path. Let represent the execution efficiency score of the r-th feasible planning path. This represents the step redundancy score for the r-th feasible planning path. This represents the resource cost score of the r-th feasible planning path. This represents the risk level score of the r-th feasible planning path. This represents the constraint satisfaction score for the r-th feasible programming path. Represents the path embryo kernel mapping matrix. This represents the aggregate signature mapping matrix of the chain trace profile. The target intent preimage mapping matrix represents the matrix. This represents the preimage mapping matrix of the constraint conditions. This represents the preimage mapping matrix for the completion criterion. The aggregate signature vector representing the chain trace profile. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. This represents the preimage vector of the completion criterion.

[0111] S303, for all feasible planning path kernels and path attribute vector Using the potential field texture operator, the policy potential field coupling matrix is ​​obtained. After this processing, the proximity and competition relationships between the paths can be expressed in a unified matrix form, thus forming a strategy potential field for multi-path comparison, expressed as:

[0112] ;

[0113] in, Let represent the potential coupling strength between the r-th feasible planning path and the s-th feasible planning path. Represents the policy potential coupling matrix. Represents the path-nucleus coupling matrix. Represents the path attribute coupling matrix. and Let represent the kernel representation vectors of the r-th and s-th feasible planning paths, respectively. and Let r and s represent the attribute vectors of the r-th and s-th feasible planning paths, respectively.

[0114] S304, for path attribute vectors Coupled matrix with strategy potential field Using the path potential energy summation operator, the path potential energy of the r-th feasible planning path is obtained. and path hierarchy tags After this processing, different paths can be ranked according to their overall efficiency level, and further distinguished into optimal paths, acceptable paths, and inefficient paths, as expressed in the following expression:

[0115] ;

[0116] in, Let r represent the path potential energy of the r-th feasible planning path. The hierarchical label represents the r-th feasible planning path. Indicates the execution efficiency weight. Indicates the redundancy weight of the steps. Indicates the resource cost weight. Indicates the risk level weight. Indicates the constraint satisfaction weight. Indicates the potential coupling weight. Let represent the execution efficiency score of the r-th feasible planning path. This represents the step redundancy score for the r-th feasible planning path. This represents the resource cost score of the r-th feasible planning path. This represents the risk level score of the r-th feasible planning path. This represents the constraint satisfaction score for the r-th feasible programming path. Let represent the potential coupling strength between the r-th feasible planning path and the s-th feasible planning path. and The threshold for hierarchical discrimination is represented, and it satisfies the following conditions: , The hierarchical discriminant function is defined when... Output the optimal path label when When the output accepts path labels, Inefficient path labels are output at times.

[0117] In the specific implementation of the above embodiments, step S4 includes the following steps:

[0118] S401, regarding the original image of the plan Ω-strategy potential field of chain track profile Using the facet convergence operator, the five-dimensional evaluation facet vector corresponding to the elasticity evaluation scale is obtained. After this processing, the degree of task objective matching, the completeness of the thought process, the rationality of tool usage, the potential energy level of the path, and the satisfaction of business constraints can all be uniformly mapped to the same quantitative space, expressed as:

[0119] ;

[0120] in, Represents the five-dimensional evaluation facet vector. Indicates the degree of matching between task objectives. Indicates the completeness of the thought process chain. Indicates the appropriateness of tool usage. Indicates the path potential energy level. Indicates the status of business constraint satisfaction. This represents the preimage mapping matrix for planning. Represents the chain track profile mapping matrix. The tool coupling mapping matrix is ​​represented. Represents the strategy potential mapping matrix. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. This represents the preimage vector of the completion criterion. The aggregate signature vector representing the chain trace profile. Indicates the relationship between the i-th semantic anchor and the th... The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool. Indicates the first The path potential energy of a feasible planning path, Let represent the attribute vector of the r-th feasible planning path, N represent the total number of semantic anchors, M represent the total number of candidate tools, and R represent the total number of feasible planning paths. This represents the normalization function.

[0121] S402, evaluation of five-dimensional faceted vectors Strategic potential field Using the task structure parameters and the elastic weight operator, a dynamic weight vector is obtained. After this processing, the evaluation weights can be adaptively adjusted according to the task complexity and the diversity of planning paths, thus making the comprehensive score result closer to the actual application scenario. The expression is:

[0122] ;

[0123] in, Represents a dynamic weight vector. , , , and The weights correspond to five evaluation dimensions: the degree of matching between task objectives, the completeness of the thought process, the rationality of tool usage, the level of path potential, and the satisfaction of business constraints. Represents the basic weight vector. This represents the task complexity adjustment factor. Indicates the path diversity adjustment coefficient. This represents the direction vector for complexity weighting. Let N represent the diversity weighting direction vector, M represent the total number of semantic anchors, L represent the total number of candidate tools, and K represent the total number of task planning steps. Indicates the first The potential coupling strength between the s-th feasible planning path and the s-th feasible planning path. This represents the normalized weight function.

[0124] S403, evaluation of five-dimensional faceted vectors and dynamic weight vector The comprehensive score is obtained by using the elasticity evaluation scale calculation operator. And the diagnostic vector z, thus forming the elasticity evaluation scale result. After this processing, the large model's task planning capability can be output with a unified score and locatable deviation information, expressed as:

[0125] ;

[0126] in, This indicates the results of the elasticity evaluation scale. This indicates the overall score. Represents the diagnostic vector. , , , and These correspond to the weighted deviation values ​​of the five evaluation facets. This represents the baseline threshold vector for the five evaluation facets. This indicates element-wise multiplication. Represents a dynamic weight vector. This represents the five-dimensional evaluation facet vector.

[0127] To demonstrate the effectiveness of this method, a public dataset was used for validation. Specifically, the publicly available dataset TRAJECT-Bench was selected for experimental validation. This dataset covers parallel and sequential tool invocation tasks and includes user queries of varying difficulty levels, effectively reflecting task planning and tool invocation scenarios under complex business processes. In the experiments, several representative models were selected to generate task planning results, and evaluations were conducted using standard answer precise matching, tool invocation legality, result-oriented evaluation, and the method of this invention. Simultaneously, human expert scoring was introduced as a reference standard to compare the consistency between different evaluation methods and human judgment, the ability to identify reasonable alternative paths, and the ability to locate procedural errors. Experiments show that this method outperforms the comparative methods in terms of correlation with human scoring, reasonable path identification rate, and procedural error identification rate, indicating that this method can more accurately reflect the true quality of complex task planning while maintaining evaluation consistency. This public benchmark itself is designed for trajectory-level tool use and includes trajectory-level indicators such as tool selection, parameter correctness, and dependency order, making it particularly suitable as a validation dataset for this method.

[0128] Table 1: Comparison of overall effectiveness of different evaluation methods

[0129] method Human consistency Reasonable path recognition rate (%) Process error identification rate (%) Result discrimination accuracy (%) stability Exact match 0.54 32.8 29.6 74.3 0.87 Validity verification 0.61 45.7 53.4 78.6 0.89 Results Evaluation 0.68 57.9 42.1 82.4 0.91 This method 0.91 88.6 89.3 91.7 0.96

[0130] Table 2: Evaluation performance of this method under different task types

[0131] Task type Difficulty Preimage matching (%) Chain trace complete (%) Path potential energy (%) Overall score Human consistency Parallel tasks Simple 92.4 90.8 89.7 91.3 0.93 Parallel tasks difficulty 88.1 86.7 84.9 86.8 0.89 Sequential tasks Simple 93.6 92.1 91.2 92.4 0.94 Sequential tasks difficulty 89.5 87.9 85.8 87.9 0.90

[0132] Table 3: Ablation results of this method

[0133] set up Human consistency Reasonable path recognition rate (%) Process error identification rate (%) Overall score Remove chain traces 0.78 71.2 63.5 76.4 Castration field 0.81 69.8 78.6 79.1 Elasticity 0.84 80.5 81.7 83.2 Complete method 0.91 88.6 89.3 91.7

[0134] Experimental results show that our proposed method significantly outperforms the comparative methods in overall performance. Taking Table 1 as an example, the human consistency score of our method reaches 0.91, significantly higher than 0.54 for exact matching, 0.61 for validity verification, and 0.68 for result evaluation. This indicates that the evaluation results of our method are closer to human judgment and better reflect the quality of real-world task planning. Regarding the reasonable path identification rate, our method reaches 88.6%, which is 55.8, 42.9, and 30.7 higher than exact matching, validity verification, and result evaluation, respectively. This demonstrates that our method can more effectively identify multiple reasonable planning paths for the same task objective, overcoming the problem of traditional methods over-reliance on a single standard answer. In terms of process error identification rate, our method reaches 89.3%, while exact matching, validity verification, and result evaluation only reach 29.6, 53.4, and 42.1, respectively. This indicates that our method has a stronger ability to identify process problems such as missing steps, inappropriate tool selection, and incorrect calling order. In terms of result discrimination accuracy, this invention achieves 91.7%, which is higher than the other three methods' 74.3, 78.6, and 82.4, indicating that this method can not only analyze the process but also maintain a high ability to judge the final result. The stability index reaches 0.96, which is also better than other methods, indicating that this method has better evaluation stability under different task samples.

[0135] As shown in Table 2, this method maintains a high performance across different task types and difficulty levels. In parallel tasks, the overall score is 91.3 for simple tasks and 86.8 for difficult tasks, a decrease of 4.5. In sequential tasks, the overall score is 92.4 for simple tasks and 87.9 for difficult tasks, with a similarly small decrease, indicating that this method remains adaptable to increasing task complexity. Meanwhile, the human consistency scores for sequential tasks reach 0.94 and 0.90, slightly higher than the 0.93 and 0.89 for parallel tasks, indicating that this method can more stably complete process evaluation in scenarios with clear dependencies. The ablation results in Table 3 further demonstrate the necessity of each technique. After removing the chain trace module, the process error recognition rate decreased from 89.3 to 63.5, a decrease of 25.8%, indicating that the chain of thought unfolding plays an important role in process analysis. After removing the potential field module, the reasonable path recognition rate decreased from 88.6 to 69.8, a decrease of 18.8%, indicating that the multi-path strategy comparison mechanism is crucial for identifying reasonable alternatives. After removing the elasticity module, the overall score decreased from 91.7 to 83.2, a drop of 8.5, indicating that the comprehensive quantification mechanism also significantly contributed to the overall evaluation effect. The above data shows that this method not only improves the consistency between evaluation results and human judgment, but also better handles multi-path planning and process quality analysis in complex tasks, verifying the effectiveness of this method.

[0136] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of patent protection of the present invention.

Claims

1. A task planning evaluation method based on thought chain and strategy optimization, characterized in that, Includes the following steps: S1, Construct the planning archetype for task planning evaluation; S2, based on the original planning image, expands the thought chain of the task planning process of the large model to generate the corresponding chain trace profile; S3, construct the policy potential field corresponding to the chain trace profile; S4, based on the planning preimage, chain trace profile and policy potential field, constructs an elastic evaluation scale to uniformly quantify the planning capability of large model tasks.

2. The task planning and evaluation method based on thought chain and strategy optimization according to claim 1, characterized in that, Step S1 includes the following steps: S101 performs joint parsing of user task text, business context, and rule information, and compresses the target intent, constraints, task objects, and business rules into a unified semantic anchor representation, expressed as: ； in, Indicates the first A semantic anchor vector, Indicates the first A target intent representation vector. Indicates the first Each constraint condition is represented by a vector. Indicates the first Each task object is represented by a vector. Indicates the first Each business rule is represented by a vector. , , and These represent the mapping matrices corresponding to the target intent, constraints, task objects, and business rules, respectively. Represents a nonlinear transformation function. Indicates the first The significance coefficient of each semantic anchor point This represents the significance adjustment coefficient, where 1 indicates that the coefficient is equal to the average value of the coefficient. All one vectors of the same dimension This indicates element-wise multiplication; S102 calculates the adaptation relationship between semantic anchors and the candidate tool set, and forms a stable coupling distribution under the combined effect of constraints and tool capability boundaries. The expression is as follows: ； in, Indicates the first The semantic anchor point and the first The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool, where B represents the boundary mapping matrix between the semantic space and the tool space. Indicates the first The semantic anchor point and the first Risk distance between candidate tools Indicates the first The semantic anchor point and the first The degree of constraint conflict among candidate tools Indicates the first The semantic anchor point and the first The capability matching gain among the candidate instruments is given by α (risk distance adjustment coefficient), β (constraint conflict adjustment coefficient), γ (capability matching adjustment coefficient), and M (total number of candidate instruments). This represents the transpose of the i-th semantic anchor vector. The vector representing the capability of the k-th candidate tool. This represents the risk distance between the i-th semantic anchor and the k-th candidate tool. This indicates the degree of constraint conflict between the i-th semantic anchor and the k-th candidate tool. This represents the capability matching gain between the i-th semantic anchor and the k-th candidate tool. This represents the operation of the natural exponential function; S103 identifies key execution nodes in the task execution chain. It quantifies the importance of computational nodes by introducing completion criteria, expressed as follows: ； in, Indicates the first The importance score of each candidate execution node, where N represents the total number of semantic anchors. Indicates the first Tool index mapping function corresponding to each candidate execution node Indicates the first Task weights for each semantic anchor point Indicates the causal enhancement coefficient. Indicates the first The causal contribution strength of each candidate execution node in the execution chain Indicates the enhancement coefficient of the completion criterion. Indicates the first Consistency score between each candidate execution node and the completion criterion; S104 unifies and aggregates the target intent, constraints, tool boundaries, key execution nodes, and completion criteria to form a planning preimage for subsequent evaluation, expressed as: ； in, Indicates the original image of the plan. This represents the aggregated target intent preimage vector. Let X represent the preimage vector of the aggregated constraint conditions, where X represents the vector formed by the total coupling strength. The tool boundary coupling matrix is ​​formed. Represents the set of critical execution nodes. This represents the preimage vector of the completion criterion. Indicates the first The weight of each semantic anchor in the aggregation of target intent Indicates the first The weights of each semantic anchor point in the constraint aggregation, where N represents the total number of semantic anchor points, M represents the total number of candidate tools, and P represents the total number of candidate execution nodes. Indicates the first The structural representation vector of each candidate execution node Represents the normalization function. This represents the importance score of all nodes. The constructed score vector, This indicates that the highest score is selected. One candidate execution node.

3. The task planning and evaluation method based on thought chain and strategy optimization according to claim 1 or 2, characterized in that: In step S1, the planning preimage is used to extract the target intent, constraints, tool boundaries, key execution nodes and completion criteria in the user task in a structured manner, and form an evaluation benchmark consistent with the actual business needs.

4. The task planning and evaluation method based on thought chain and strategy optimization according to claim 2, characterized in that, Step S2 includes the following steps: S201, regarding the original plan The task planning step sequence output by the large model Using the preimage projection expansion operator, we obtain the preimage-aligned thought step representation, expressed as: ； in, Let represent the representation vector of the u-th original task planning step, and L represent the total number of task planning steps. Let the vector represent the thought step of the u-th preimage alignment. This represents the original planning step mapping matrix. The target intent preimage mapping matrix represents the matrix. This represents the preimage mapping matrix of the constraint conditions. Represents the tool boundary mapping matrix. This represents the mapping matrix of key execution nodes. This represents the preimage mapping matrix for the completion criterion. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. Indicates the first The semantic anchor point and the first The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool. Indicates the first The importance score of each candidate execution node Indicates the first The structural representation vector of each candidate execution node Indicates the first Does each candidate execution node belong to the set of critical execution nodes? The selection flag, when the node belongs to hour ,otherwise K represents the number of critical execution nodes. This represents the preimage vector of the completion criterion. This represents the explicit gain coefficient for the u-th task planning step. Represents the normalization function; S202, Preimage Alignment Thinking Step Representation Using a multi-track splitting mapping operator, we obtain track-by-track representations of the target recognition trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory, and step connection trajectory. The expression is: ； in, Let represent the trajectory sub-vector of the u-th task planning step on the z-th type trajectory, where z is the trajectory type number, z=1 represents the target recognition trajectory, z=2 represents the subtask decomposition trajectory, z=3 represents the tool selection trajectory, z=4 represents the parameter organization trajectory, and z=5 represents the step connection trajectory. This represents the step-state mapping matrix corresponding to the z-th type of trajectory. This represents the semantic anchor mapping matrix corresponding to the z-th type of trajectory. This represents the tool capability mapping matrix corresponding to the z-th type of trajectory. In the planning step of the u-th task, the first step is... The contribution weights of each semantic anchor point to the z-th type trajectory Indicates the first A semantic anchor vector, In the planning step of the u-th task, the first step is... The contribution weights of each candidate tool to the z-th type trajectory; S203, representing five categories of track division Using the sheet weaving operator, we obtain chain trace sheets that characterize the internal structure of the current planning step and the connection between the preceding and following steps. The expression is: ； in, Indicates the first The chain trace layer slice vector corresponding to each task planning step This represents the intra-layer contribution coefficient of the z-th type trajectory in the u-th task planning step. In the planning step of the u-th task, the first step is... Class trajectory and the first The weaving coefficient between class trajectories, Indicates the selection marker for the trajectory pair, when hour ,when hour , This indicates element-wise multiplication. This represents the continuation coefficient of the u-th task planning step from the previous task planning step. The vector represents the step-connection trajectory of the previous task planning step. Number 5 corresponds to the step-connection trajectory. hour, Take the zero vector; S204, for all chain trace layers And the corresponding five types of trajectory sequences, using the profile aggregation operator, yield the chain track profile Ω, expressed as: ； Where Ω represents the chain track profile, Represents the target recognition trajectory sequence. This represents the sequence of subtask decomposition trajectories. The tool indicates that a trajectory sequence has been selected. This indicates the parameter organization of the trajectory sequence. This indicates the sequence of steps connecting the trajectory. Indicates that it consists of all chain trace layers. The constructed layer sequence, The aggregate signature vector representing the chain trace profile. This represents the weight of the u-th task planning step in the profile aggregation.

5. The task planning and evaluation method based on thought chain and strategy optimization according to claim 1 or 4, characterized in that: In step S2, the chain trace profile is used to characterize the target recognition trajectory, subtask decomposition trajectory, tool selection trajectory, parameter organization trajectory and step connection trajectory formed by the model in the task solving process, and presents the key cognitive links in the planning process in a hierarchical manner.

6. The task planning and evaluation method based on thought chain and strategy optimization according to claim 4, characterized in that, Step S3 includes the following steps: S301, comparing the chain track profile Ω with the planned preimage Using the path embryo kernel extraction operator, feasible planning path embryo kernel representations around the same task objective are obtained. The expression is: ； in, Let r be the kernel representation vector of the r-th feasible planning path, where r = 1, 2, ..., R; and R represent the total number of feasible planning paths. Represents the chain trace layer-slice mapping matrix. The trace mapping matrix represents the steps connecting the trajectory. This represents the mapping matrix of key execution nodes. Let represent the assignment coefficient of the u-th task planning step to the r-th feasible planning path. This represents the chain trace layer vector corresponding to the u-th task planning step. Let represent the step connection trajectory vector for the u-th task planning step, where number 5 corresponds to the step connection trajectory. Indicates the first Does each candidate execution node belong to the set of critical execution nodes? The selection marker, Indicates the first The importance score of each candidate execution node Indicates the first The structural representation vector of each candidate execution node Represents the normalization function; S302, for feasible planning path kernels Chain trace profile aggregate signature and original plan Using the potential ripple projection operator, the attribute vector of the r-th feasible planning path is obtained. The expression is: ； in, Let represent the attribute vector of the r-th feasible planning path. Let represent the execution efficiency score of the r-th feasible planning path. This represents the step redundancy score for the r-th feasible planning path. This represents the resource cost score of the r-th feasible planning path. This represents the risk level score of the r-th feasible planning path. This represents the constraint satisfaction score for the r-th feasible planning path. Represents the path embryo kernel mapping matrix. This represents the aggregate signature mapping matrix of the chain trace profile. The target intent preimage mapping matrix represents the matrix. This represents the preimage mapping matrix of the constraint conditions. This represents the preimage mapping matrix for the completion criterion. The aggregate signature vector representing the chain trace profile. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. This represents the preimage vector of the completion criterion; S303, for all feasible planning path kernels and path attribute vector Using the potential field texture operator, the policy potential field coupling matrix is ​​obtained. The expression is: ； in, Let represent the potential coupling strength between the r-th feasible planning path and the s-th feasible planning path. Represents the policy potential coupling matrix. Represents the path-nucleus coupling matrix. Represents the path attribute coupling matrix. and Let represent the kernel representation vectors of the r-th and s-th feasible planning paths, respectively. and Let r and s represent the attribute vectors of the r-th and s-th feasible planning paths, respectively. S304, for path attribute vectors Coupled matrix with strategy potential field Using the path potential energy summation operator, the path potential energy of the r-th feasible planning path is obtained. and path hierarchy tags The expression is: ； in, Let r represent the path potential energy of the r-th feasible planning path. The hierarchical label represents the r-th feasible planning path. Indicates the execution efficiency weight. Indicates the redundancy weight of the steps. Indicates the resource cost weight. Indicates the risk level weight. The weights represent the constraints that are satisfied. Indicates the potential coupling weight. Let represent the execution efficiency score of the r-th feasible planning path. This represents the step redundancy score for the r-th feasible planning path. This represents the resource cost score of the r-th feasible planning path. This represents the risk level score of the r-th feasible planning path. This represents the constraint satisfaction score for the r-th feasible planning path. Let represent the potential coupling strength between the r-th feasible planning path and the s-th feasible planning path. and The threshold for hierarchical discrimination is represented, and it satisfies the following conditions: , The hierarchical discriminant function is defined when... Output the optimal path label when When the output accepts path labels, Inefficient path labels are output at times.

7. The task planning and evaluation method based on thought chain and strategy optimization according to claim 1 or 6, characterized in that: In step S3, the strategy potential field is used to characterize the distribution of feasible planning paths formed around the same planning preimage, and to conduct a unified analysis on the different paths in terms of execution efficiency, step redundancy, resource cost, risk level and constraint satisfaction. Based on this, a path potential energy calculation mechanism is established to measure the comprehensive efficiency level of each planning path in the process of achieving the goal.

8. The task planning and evaluation method based on thought chain and strategy optimization according to claim 6, characterized in that, Step S4 includes the following steps: S401, regarding the original image of the plan Ω-strategy potential field of chain trace profile Using the facet convergence operator, the five-dimensional evaluation facet vector corresponding to the elasticity evaluation scale is obtained. The expression is: ； in, Represents the five-dimensional evaluation facet vector. Indicates the degree of matching between task objectives. Indicates the completeness of the thought process chain. Indicates the appropriateness of tool usage. Indicates the path potential energy level. Indicates the status of business constraint satisfaction. This represents the preimage mapping matrix. Represents the chain track profile mapping matrix. The tool coupling mapping matrix is ​​represented. Represents the strategy potential mapping matrix. Represents the target intent preimage vector. Represents the preimage vector of the constraint conditions. This represents the preimage vector of the completion criterion. The aggregate signature vector representing the chain trace profile. Indicates the relationship between the i-th semantic anchor and the th... The coupling strength between candidate tools Indicates the first A vector representing the capabilities of each candidate tool. Indicates the first The path potential energy of a feasible planning path, The attribute vector representing the r-th feasible planning path; S402, evaluation of five-dimensional faceted vectors Strategic potential field Using the task structure parameters and the elastic weight operator, a dynamic weight vector is obtained. The expression is: ； in, Represents a dynamic weight vector. , , , and The weights correspond to five evaluation dimensions: the degree of matching between task objectives, the completeness of the thought process, the rationality of tool usage, the level of path potential, and the satisfaction of business constraints. Represents the basic weight vector. This represents the task complexity adjustment factor. This represents the path diversity adjustment coefficient. This represents the direction vector for complexity weighting. Represents the direction vector of diversity weighting. Indicates the first The potential coupling strength between the s-th feasible planning path and the s-th feasible planning path. This represents the normalized weight function; S403, evaluation of five-dimensional faceted vectors and dynamic weight vector The comprehensive score is obtained by using the elasticity evaluation scale calculation operator. And the diagnostic vector z, thus forming the elasticity evaluation scale result. The expression is: ； in, This indicates the results of the elasticity evaluation scale. This indicates the overall score. Represents the diagnostic vector. , , , and These correspond to the weighted deviation values ​​of the five evaluation facets. This represents the baseline threshold vector for the five evaluation facets. This indicates element-wise multiplication. Represents a dynamic weight vector. This represents the five-dimensional evaluation facet vector.

9. The task planning and evaluation method based on thought chain and strategy optimization according to claim 1 or 8, characterized in that: In step S4, the flexible evaluation scale is used to comprehensively consider the degree of matching of task objectives, the completeness of the thought chain, the rationality of tool calls, the level of path potential energy, and the satisfaction of business constraints. It also dynamically adjusts the evaluation weights in combination with the complexity of the task and the diversity of the planned path to form a comprehensive score result.

Images