A method, system, and application for automatic generation and optimization of hybrid workflows driven by a large model.

Abstract

This invention discloses a large-model-driven method for the automatic generation and optimization of hybrid workflows, comprising: receiving user task requirements and constructing an initial workflow; establishing a space representing the structural and computational characteristics of the workflow, and iteratively optimizing the workflow within this space; during the optimization process, analyzing historical feedback using a large model and generating improvement strategies, thereby generating candidate workflows that integrate semantic reasoning units and deterministic computation units; performing execution tests on the candidate workflows, and filtering and updating them through multi-dimensional performance evaluation until a heterogeneous workflow meeting preset conditions is formed. This invention also discloses a system for implementing the above optimization method, which has broad application value. This invention can automatically construct heterogeneous workflows for complex tasks, significantly improving task processing accuracy while greatly reducing inference costs and execution time, achieving optimal resource utilization.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and automation technology, and relates to a method, system and application for automatic construction and multi-objective optimization of hybrid workflows using large language models (LLM) and evolutionary algorithms. Background Technology

[0002] With the rapid development of information technology, facing increasingly complex business needs and multimodal data processing tasks, breaking down complex tasks into workflows composed of multiple ordered steps has become the mainstream paradigm for achieving automated processing. In the traditional model, workflow construction typically relies on domain experts manually orchestrating based on experience, or on filling in structures using fixed templates. These methods are highly dependent on human knowledge, have high development and maintenance costs, and the resulting process structures lack flexibility, making it difficult to adapt to constantly changing user needs and complex unstructured task scenarios.

[0003] In recent years, large language models have demonstrated significant advantages in task understanding and semantic reasoning, prompting researchers to explore technical approaches for automatically generating workflows. However, existing methods mostly employ a one-time generation paradigm, where the system directly outputs a complete workflow structure or executable script based on user instructions. These methods lack a closed-loop iterative mechanism based on execution feedback. If deviations occur in the generated process's logical structure, parameter settings, or execution stages, the system often fails to automatically identify and correct them, making it difficult to effectively guarantee the reliability and stability of complex tasks.

[0004] Besides lacking adaptive optimization capabilities, existing workflow generation technologies also have significant limitations in how they handle execution units. Current optimization methods typically use a predefined toolkit as the search space, mainly arranging and combining existing component connections, focusing on topology tuning. This type of stock-based search cannot automatically create new processing units based on task requirements, forcing the system to rely on generic large-model inference interfaces when encountering unseen tasks. Over-reliance on large models to handle the entire process not only leads to unstable outputs in scenarios such as precise calculations, logic verification, and large-scale data transformations, but also incurs high inference costs.

[0005] More critically, existing technologies generally lack the ability to build customized execution units, failing to automatically generate reusable deterministic computation modules based on task requirements. For example, when handling specific mathematical calculations, data cleaning, or structured transformation tasks, the system cannot automatically generate corresponding efficient code and encapsulate it into independently runnable units. The combination of predefined components and general large model interfaces significantly limits the overall workflow in terms of accuracy, execution efficiency, and resource utilization. Therefore, current automated workflow generation technologies are insufficient in terms of structural flexibility, feedback-driven adaptive optimization capabilities, and the generativeness of execution units. There is an urgent need for a hybrid workflow automatic generation and optimization method that integrates semantic reasoning capabilities and deterministic computation capabilities, and possesses an iterative optimization mechanism, to address the comprehensive requirements of reliability, performance, and cost in complex task scenarios. Summary of the Invention

[0006] To address the problems of existing workflow generation methods, such as reliance on manual design, lack of adaptive optimization capabilities, and inability to automatically generate deterministic execution units based on task requirements, leading to instability and high costs in complex task execution, this invention aims to provide a hybrid workflow automatic generation method and system that integrates semantic reasoning capabilities and deterministic computation capabilities. Starting from the user's task description and building upon an initial workflow, this method automatically generates a hybrid workflow structure composed of semantic reasoning units and deterministic computation units through a feedback-driven iterative evolution mechanism. This effectively improves workflow execution stability, task completion accuracy, and overall resource utilization efficiency.

[0007] The large-model-driven method for automatic generation and optimization of hybrid workflows proposed in this invention includes the following steps:

[0008] Step 1: The controller receives the user's task requirements, and the system constructs an initial workflow. The initial workflow is a directed acyclic graph, whose node set contains only one semantic reasoning unit, and whose edge set contains only the structural edges that pass the user input to the semantic reasoning unit.

[0009] Step 2: Run the initial workflow to execute the target task and obtain performance information characterizing its functional performance, including task completion rate, response time and inference cost;

[0010] Step 3: Construct an elite profile based on the structural and computational features of the workflow. The structural features represent the scale and topological complexity of the workflow, and the computational features represent the combination and distribution of different types of processing units in the workflow. Map the initial workflow and its performance information to feature space coordinates. If there is no archived individual at that coordinate position, store the initial workflow and its performance information in the elite profile.

[0011] Step 4: Perform iterative iterations based on the elite portfolio. Each iteration includes: selecting a parent workflow from the elite portfolio; using a large model-driven evolutionary generator to generate an improvement strategy in natural language form based on the historical feedback of the parent workflow; performing structural mutation on the parent workflow according to the improvement strategy to generate a new workflow; executing the newly generated workflow to obtain performance feedback; performing multi-objective evaluation on the newly generated workflow based on the performance feedback; and updating the elite portfolio based on the evaluation results.

[0012] Step 5: When the preset termination condition is met, select the workflow with the highest comprehensive score from the elite files as the optimal workflow output. The optimal workflow is a heterogeneous workflow that integrates semantic reasoning units and deterministic computing units.

[0013] The nodes of the directed acyclic graph described in this invention are composed of a mixture of semantic reasoning units and deterministic computing units; wherein, the semantic reasoning units process unstructured reasoning tasks based on a large language model, and the deterministic computing units process logical computing tasks based on executable code; the directed edges in the directed acyclic graph represent data flow or control dependencies between the execution units.

[0014] The multi-objective evaluation described in this invention employs a scoring formula that includes execution performance, resource cost, and response time. The scoring formula is as follows:

[0015] ,

[0016] in, The task accuracy score is calculated based on the number of correct results / total number of tasks. To calculate the inference cost score based on the number of tokens consumed, Scoring is based on response time. These are the weighting coefficients.

[0017] The present invention updates the elite archive by: mapping the structural features and computational features of the newly generated workflow to feature coordinates in the archive space; if there is no individual at the coordinate position or the newly generated workflow is superior to the existing individual at the coordinate position, then the newly generated workflow and its performance information are stored in the elite archive.

[0018] Step 4 of this invention, selecting a parent workflow from the elite archive, specifically includes: selecting one or more stored workflows from the elite archive as parent workflows according to a preset sampling strategy. The preset sampling strategy specifically includes: sorting the workflows stored in the elite archive based on performance indicators, selecting the workflow with the best performance as the basic parent; locating the position of the basic parent in the feature space using a feature coordinate system, retrieving feature units adjacent to its feature coordinates, and selecting workflows from the adjacent feature units as supplementary parents.

[0019] The process of generating an improvement strategy using the evolutionary generator of this invention includes: the evolutionary generator reads the historical execution results and state distribution of the parent workflow and elite files, uses the reasoning ability of a large language model for reflection and diagnosis, and generates optimization suggestions in natural language form as the improvement strategy. The step of generating a new workflow based on the improvement strategy specifically includes: the evolutionary generator performing structural mutation on the parent workflow according to the improvement strategy. The structural mutation operations include adding execution units, removing execution units, changing the type of execution units, or adjusting the data dependencies between execution units, and the total number of nodes in the workflow is constrained to a preset range of 1 to 8.

[0020] Preferably, the generation of the new workflow further includes a unit logic generation step: when the structural variation involves the construction or modification of a semantic reasoning unit, a natural language prompt word template that drives the unit to run is generated by the large model; when the structural variation involves the construction or modification of a deterministic computing unit, the executable program code of the unit is generated by the large model.

[0021] Based on the above, this invention also proposes a large-model-driven hybrid workflow automatic generation and optimization system, comprising:

[0022] The initialization module is used to receive task requirements and obtain the initial workflow.

[0023] The archive management module is used to build and maintain elite archives based on structural and computational features. The structural features characterize the scale and topological complexity of the workflow, while the computational features characterize the combination and distribution of different types of processing units in the workflow.

[0024] The evolutionary iteration module is used to select parent workflows from the elite files, drive the large model to reflect on and analyze the historical feedback of the parent workflows to generate improvement strategies, generate new workflows with executable code based on the strategies, perform execution tests on the new workflows to obtain performance feedback, and perform multi-objective evaluation on the new workflows based on the performance feedback, and update the elite files based on the evaluation results.

[0025] The output module is used to select the workflow with the highest comprehensive score from the elite files as the optimal workflow output when the preset termination conditions are met.

[0026] The various modules in the system work together to achieve automatic generation, continuous optimization, and performance balancing of workflows.

[0027] Based on the above, this invention also proposes applications of the above-mentioned methods and systems in complex logical reasoning tasks (such as solving mathematical problems, automatic code generation, data cleaning, multi-step logical reasoning, and structured transformation).

[0028] The large-model-driven automatic generation and optimization method and system for hybrid workflow disclosed in this invention have the following beneficial effects:

[0029] 1. This invention, through a closed-loop iterative mechanism of "reflection-generation-execution-evaluation", improves the final task resolution rate by about 100% compared to the initial state, achieving a continuous leap in performance;

[0030] 2. The hybrid heterogeneous workflow generated by this invention effectively overcomes the illusion of a large model, improving task accuracy by 11.09 percentage points;

[0031] 3. This invention uses deterministic computing units to replace reasoning steps, which reduces reasoning costs by about 9 times and shortens task execution time by about 2 times. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the system architecture and initial workflow of the present invention;

[0034] Figure 2 The experimental results of the iterative optimization strategy in this embodiment of the invention are shown in the figure (curves showing the change in the resolution rate of the validation set and the test set as the number of iterations increases).

[0035] Figure 3 This is an experimental result diagram showing the effectiveness of the hybrid workflow generation method of this invention (comparison data between this invention and advanced methods in the field on multiple dimensions). Detailed Implementation

[0036] The invention will be further described in detail below with reference to the specific embodiments and accompanying drawings. Except for the contents specifically mentioned below, the processes, conditions, and experimental methods for implementing the invention are all common knowledge and general knowledge in the art, and the invention does not have any particular limitations.

[0037] This invention provides a large model-driven hybrid workflow automatic generation and optimization method. This method integrates the semantic reasoning ability of a large language model with the deterministic computational ability of code generation. Through a closed-loop iterative mechanism of "reflection-generation-execution-evaluation", it solves the problems of workflow construction relying on manual labor, unstable execution, and high cost in complex tasks.

[0038] This invention can handle complex logical reasoning and unstructured task scenarios. Through a large model-driven evolutionary iteration approach, it automatically and efficiently generates hybrid heterogeneous workflows tailored to specific task requirements. This overcomes the limitations of traditional manual orchestration or direct reasoning with a single model when facing increasing task complexity, such as frequent illusions, high reasoning costs, and insufficient flexibility. It can effectively improve the processing quality and execution efficiency of complex tasks while ensuring high accuracy, significantly reducing overall reasoning costs and response time, and also has good adaptability and robustness.

[0039] This invention includes the following steps:

[0040] Step 1: Obtain the initial workflow

[0041] The controller receives user task requests, and the system builds the initial workflow. (Right now Figure 1 The initial workflow (in the context of the initial workflow). Defined as a set of nodes and edge set The directed acyclic graph formed, in which the node set It contains only one general semantic reasoning unit. edge set It only contains structural edges that connect to the input.

[0042] Step 2: Test benchmark performance

[0043] Controller runs initial workflow It records the execution results on the target task, including performance information such as task completion rate, response time, and inference cost, and represents this performance information as... ,in An evaluation function defined for the system, used to provide an initial reference for subsequent iterations.

[0044] Step 3: Construct elite profiles

[0045] Controller calculation workflow The structural and computational characteristics are used as the horizontal and vertical axes, respectively, to... Mapped to a specific location in a two-dimensional feature space The x-axis is... The vertical axis represents the scale and topology complexity of the workflow. This characterizes the type combination and distribution of processing units. If the coordinate position is not yet occupied, the current workflow and corresponding performance information will be displayed. Stored in the elite archives.

[0046] Step 4: Iterative Evolution

[0047] The system continues to iterate and optimize until the termination condition is met. In each iteration, the system samples the parent workflow from the elite archives based on performance ranking and a feature neighborhood search strategy. The execution records and archive status are used as context input to the large language model, which then generates an improvement strategy for the natural language form of the parent. The evolutionary generator mutates the workflow structure according to this strategy, including adding nodes, deleting nodes, replacing node types, or adjusting dependencies. During this process, the total number of nodes in the workflow is constrained to a preset range of 1 to 8 to prevent excessive structural complexity and ensure execution efficiency. When mutation involves deterministic logic, the large model simultaneously generates executable code, enabling the new nodes to have stable computational functions, thereby forming new candidate workflows. The system then... Perform execution tests to obtain performance. The overall performance of candidate workflows is evaluated based on a multi-objective scoring function, the formula of which is as follows:

[0048]

[0049] in, The task accuracy score is calculated based on the number of correct results / total number of tasks. The inference cost score is calculated based on the number of tokens consumed. Scoring is based on response time. , where is the weighting coefficient, which can be preset with weights. To balance accuracy and resource efficiency;

[0050] when When an individual surpasses existing archived individuals in its characteristic space coordinates, it is written into the elite archive, enabling continuous evolution of its structure and function.

[0051] Step 5: Output the optimal workflow

[0052] Once the iteration limit is reached, the system selects the workflow with the highest overall score from the elite portfolio. The resulting heterogeneous workflow, as the final output, includes both semantic reasoning units and automatically generated deterministic computation units, enabling efficient processing of task requirements.

[0053] The specific algorithm is as follows:

[0054]

[0055] This invention can handle complex logical reasoning and unstructured task scenarios. Through a large model-driven evolutionary iteration approach, it automatically and efficiently generates hybrid heterogeneous workflows tailored to specific task requirements. This overcomes the limitations of traditional manual orchestration or direct reasoning with a single model when facing increasing task complexity, such as frequent illusions, high reasoning costs, and insufficient flexibility. It can effectively improve the processing quality and execution efficiency of complex tasks while ensuring high accuracy, significantly reducing overall reasoning costs and response time, and also has good adaptability and robustness.

[0056] Example

[0057] The method described in this invention focuses on complex task scenarios. As a specific embodiment, the process is as follows:

[0058] This invention uses the MATH dataset as experimental cases. This dataset contains highly complex mathematical competition problems and serves as an authoritative benchmark for evaluating the logical reasoning capabilities of large models. To verify the invention's ability to handle highly difficult tasks, the experiment specifically selected 617 problems at difficulty level 5 to form the experimental dataset, covering various types such as algebra, probability theory, number theory, and calculus. These problems typically require a combination of multi-step reasoning and precise calculation to solve.

[0059] Figure 1 The system architecture and initial workflow of the present invention are shown. The initial workflow is set as a basic direct prompt mode. Figure 2 The curves show the change in the resolution rate of the validation set and the test set as the number of iterations increases. Figure 3 The presentation showcases comparative data between this invention and state-of-the-art methods across multiple dimensions. The experimental cases cover several key stages, from semantic understanding and logical planning to code execution; the collaboration and structural optimization among these stages are crucial for the accuracy and resource efficiency of the entire system.

[0060] The automatically generated and optimized code in this embodiment is deployed in a conventional server environment and configured with a Python 3.9 runtime environment. The core large language model of the system is DeepSeek-Chat. The semantic reasoning and code generation capabilities of the model are obtained by calling the API interface provided by the DeepSeek open platform. The maximum number of iteration rounds is set to 40 rounds. This value is based on the previous convergence experiments. The experimental results show that the system has achieved complete convergence after 40 iterations. At this point, the performance index reaches a stable state, and further increasing the number of iteration rounds does not bring significant benefits. Therefore, this setting can balance computational efficiency and optimization effect.

[0061] To accurately evaluate the generalization performance of the generated workflow, the experiment divided the dataset into a validation set and a test set in a 1:4 ratio. The validation set was used to calculate performance feedback and drive the evolution direction during the closed-loop iteration process of "reflection-generation-execution-evaluation"; the test set was used to perform a final objective performance evaluation of the best-performing workflow output by the system after the iteration.

[0062] By recording the performance changes of this invention during 40 iterations, the effectiveness of the evolutionary strategy was analyzed, and the experimental results are as follows: Figure 2 As shown in the figure, the experimental results demonstrate a significant stepwise increase in task resolution rate with increasing iteration rounds. The initial workflow achieved a resolution rate of approximately 35% on the test set, while after optimization, the final workflow achieved a resolution rate exceeding 70%. Specifically, the system introduced computational units through structural exploration in rounds 1-10, and introduced a verification mechanism through logic enhancement in rounds 11-30, demonstrating the effectiveness of the iterative optimization mechanism proposed in this invention in improving model robustness.

[0063] To verify the overall performance of the hybrid workflow generated by this invention, the optimal workflow generated by this invention was compared with the state-of-the-art method (AFlow) in the current field under the same experimental settings. Accuracy, inference cost, and task execution time were recorded in detail for in-depth comparative analysis. The experimental results are as follows: Figure 3 As shown in the figure, the experimental results demonstrate that this invention significantly outperforms existing technologies in all key indicators. Specifically, the heterogeneous workflow solution proposed in this invention achieves higher accuracy and lower resource consumption, improving task accuracy by 11.09 percentage points, reducing inference costs by approximately 9 times, and shortening task execution time by approximately 2 times. This provides strong empirical support for the effectiveness of the proposed hybrid structure, proving its significant potential for cost reduction and efficiency improvement.

[0064] This embodiment, supported by a closed-loop iterative mechanism and a hybrid architecture, achieves a significant leap in performance for handling complex tasks through the efficient collaboration of semantic reasoning and deterministic computation, with the solution rate approximately doubling compared to the initial state. Experimental results on real-world mathematical competition datasets fully demonstrate the significant superiority of the proposed method in terms of accuracy, cost control, and execution efficiency. In particular, mathematical problem-solving, as a typical representative of "strong logic, long chains, and high precision" tasks, shares a high degree of isomorphism in its requirements for rigorous reasoning and computational accuracy with high-end industrial scenarios such as automated programming, financial risk control, and industrial process automation.

[0065] Therefore, the successful implementation of this invention not only proves the effectiveness of the hybrid architecture on academic benchmarks, but also demonstrates its generalization potential for migration to a wide range of industrial fields, providing a universal technical paradigm for large models to solve complex unstructured problems in various vertical industries.

[0066] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0067] As used in this invention, the term "comprising" is an open-ended expression, meaning it includes the contents specified in this invention but does not exclude other aspects.

[0068] As used in this invention, the term "and / or" includes any one or more of the related listed items and all combinations thereof.

[0069] The scope of protection of this invention is not limited to the above embodiments. Any variations and advantages that can be conceived by those skilled in the art without departing from the spirit and scope of the inventive concept are included in this invention and are protected by the appended claims.

Claims

1. A method for automatic generation and optimization of hybrid workflows driven by a large model, characterized in that, Includes the following steps: Step 1: The controller receives the user's task requirements, and the system constructs an initial workflow. The initial workflow is a directed acyclic graph, whose node set contains only one semantic reasoning unit, and whose edge set contains only the structural edges that pass the user input to the semantic reasoning unit. Step 2: Run the initial workflow to execute the target task and obtain performance information characterizing its functional performance, including task completion rate, response time and inference cost; Step 3: Construct an elite profile based on the structural and computational features of the workflow. The structural features represent the scale and topological complexity of the workflow, and the computational features represent the combination and distribution of different types of processing units in the workflow. Map the initial workflow and its performance information to feature space coordinates. If there is no archived individual at that coordinate position, store the initial workflow and its performance information in the elite profile. Step 4: Perform iterative iterations based on the elite portfolio. Each iteration includes: selecting a parent workflow from the elite portfolio; using a large model-driven evolutionary generator to generate an improvement strategy in natural language form based on the historical feedback of the parent workflow; performing structural mutation on the parent workflow according to the improvement strategy to generate a new workflow; executing the newly generated workflow to obtain performance feedback; performing multi-objective evaluation on the newly generated workflow based on the performance feedback; and updating the elite portfolio based on the evaluation results. Step 5: When the preset termination condition is met, select the workflow with the highest comprehensive score from the elite files as the optimal workflow output. The optimal workflow is a heterogeneous workflow that integrates semantic reasoning units and deterministic computing units.

2. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 1, characterized in that, The nodes of the directed acyclic graph are composed of a mixture of semantic reasoning units and deterministic computing units; wherein, the semantic reasoning units process unstructured reasoning tasks based on a large language model, and the deterministic computing units process logical computing tasks based on executable code; the directed edges in the directed acyclic graph represent data flow or control dependencies between the execution units.

3. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 1, characterized in that, The multi-objective evaluation employs a scoring formula that includes execution performance, resource cost, and response time. The scoring formula is as follows: ， in, The task accuracy score is calculated based on the number of correct results / total number of tasks. To calculate the inference cost score based on the number of tokens consumed, Scoring is based on response time. These are the weighting coefficients.

4. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 1, characterized in that, Updating the elite archives includes: mapping the structural and computational features of the newly generated workflow to feature coordinates in the archive space; if there is no individual at the coordinate location or the newly generated workflow is superior to the existing individual at the coordinate location, then the newly generated workflow and its performance information are stored in the elite archives.

5. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 1, characterized in that, Step four, selecting parent workflows from the elite archives, specifically includes: selecting one or more stored workflows from the elite archives as parent workflows according to a preset sampling strategy.

6. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 5, characterized in that, The preset sampling strategy specifically includes: sorting the workflows stored in the elite files according to performance indicators, selecting the workflow with the best performance as the basic parent; using the feature coordinate system to locate the position of the basic parent in the feature space, retrieving the feature units adjacent to its feature coordinates, and selecting the workflow from the adjacent feature units as the supplementary parent.

7. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 1, characterized in that, The process of generating improvement strategies by the evolution generator includes: the evolution generator reads the historical execution results of the parent workflow and the state distribution of the elite files, uses the reasoning ability of the large language model to perform reflection and diagnosis, and generates optimization suggestions in natural language form as the improvement strategies.

8. The method for automatic generation and optimization of hybrid workflow driven by a large model according to claim 7, characterized in that, The process of generating a new workflow based on the improvement strategy specifically includes: the evolution generator performing structural mutation on the parent workflow based on the improvement strategy. The structural mutation operation includes adding execution units, removing execution units, changing the type of execution units, or adjusting the data dependencies between execution units, and the total number of nodes in the workflow is constrained to a preset range of 1 to 8. Preferably, the generation of the new workflow further includes a unit logic generation step: when the structural variation involves the construction or modification of a semantic reasoning unit, a natural language prompt word template that drives the unit to run is generated by the large model; when the structural variation involves the construction or modification of a deterministic computing unit, the executable program code of the unit is generated by the large model.

9. A large-model-driven hybrid workflow automatic generation and optimization system, characterized in that, include: The initialization module is used to receive user task requirements and build the initial workflow. The document management module is used to build and maintain elite documents based on the structural and computational characteristics of the workflow. The structural characteristics represent the scale and topological complexity of the workflow, and the computational characteristics represent the combination and distribution of different types of processing units in the workflow. The evolutionary iteration module is used to select parent workflows from the elite files, drive the large model to reflect on and analyze the historical feedback of the parent workflows to generate improvement strategies, generate new workflows containing executable code based on the improvement strategies, perform execution tests on the new workflows to obtain performance feedback, perform multi-objective evaluation on the new workflows based on the performance feedback, and update the elite files based on the evaluation results. The output module is used to select the workflow with the highest comprehensive score from the elite files as the optimal workflow output when the preset termination conditions are met.

10. The application of the method as described in any one of claims 1-8 or the system as described in claim 9 in complex logical reasoning tasks, wherein the complex logical reasoning tasks include handling specific mathematical calculations, data cleaning, automatic code generation, multi-step logical reasoning, or structured transformation tasks.

Images