A crowd-sensing task allocation method and system based on decoupling of prompt words and inference

Abstract

This invention relates to the field of task allocation technology, and in particular to a crowd-aware task allocation method and system based on decoupling of cue words and inference. The method includes multi-objective task modeling and candidate worker stratification based on acquired data; local selection based on cue words within a group, including structured representation of candidate worker attributes, cue-based multidimensional individual evaluation, cross-worker comparison and trade-off reasoning within the group, and structured output and robust selection mechanism; and global optimization under budget constraints based on the reasoning results, including candidate set construction and task-level profile organization, cue word-based multi-solution generation and mind tree reasoning, and selection of the globally optimal solution under budget constraints. This invention significantly improves the efficiency and scalability of task allocation in large-scale dynamic scenarios by decoupling symbolic computation and neural reasoning.

Description

Technical Field

[0001] This invention relates to the field of task allocation technology, and in particular to a crowd-sensing task allocation method and system based on the decoupling of prompt words and inference. Background Technology

[0002] In real-world crowdsourced sensing scenarios, task allocation faces multiple complex constraints: each task typically includes strict geographical boundaries (e.g., traffic flow data must be collected at specific intersections) and effective time windows (e.g., noise monitoring only during evening rush hours); participating users possess personalized movement trajectories, dynamically changing availability, varying sensor accuracy, and individual preferences regarding task rewards, energy consumption, and privacy risks. More importantly, task allocation objectives are often multi-dimensional and conflicting. The system aims to maximize spatial coverage and sensing data quality while minimizing worker movement distances and energy consumption, and ensuring fairness and user satisfaction in task allocation. This challenge of finding optimal trade-offs among multiple objectives in a highly dynamic and uncertain environment renders traditional centralized, static scheduling strategies ineffective.

[0003] To address these challenges, academia and industry have conducted extensive research and developed various technical approaches. Heuristic optimization methods, such as genetic algorithms and particle swarm optimization, approximate optimal allocation by designing specific search operators, but they are prone to getting trapped in local optima and experience a sharp increase in computational cost as the task scales up. Mechanism design methods based on auctions and game theory guide workers' self-interested behavior to align with system goals through economic incentives, but they often assume perfectly rational participants and require high real-time communication, making them difficult to handle real-world scenarios with frequent worker entry and exit and dynamically changing preferences. Data-driven prediction methods use historical trajectory data to predict the future spatiotemporal distribution of workers for pre-assignment, but their accuracy is limited by data sparsity and environmental variability, resulting in insufficient generalization ability. In recent years, allocation models based on deep reinforcement learning have gradually become a research hotspot. Through continuous trial and error learning of implicit allocation strategies by agents and the environment, they have achieved better results than traditional methods in certain simulated environments.

[0004] On the other hand, existing technologies still suffer from several common problems that hinder their application in real-world, complex scenarios. First, the modeling complexity is too high. Most methods rely on precise mathematical objective functions and constraints, such as pre-defined worker utility functions, task reward models, or carefully designed reinforcement learning reward functions. This makes model design heavily dependent on the experience of domain experts, and when system parameters change, remodeling or even retraining is required, resulting in poor adaptability. Second, the multi-objective conflict handling mechanism is rigid. Existing technologies typically use weighted summation or priority ranking to transform multiple objectives into a single objective. Weights or priorities need to be manually preset and are difficult to dynamically adjust, making it impossible to adaptively balance coverage quality, energy consumption, fairness, and other indicators based on environmental conditions during runtime. Third, interpretability is lacking. Task allocation models based on deep neural networks are like "black boxes" in their decision-making process. System administrators find it difficult to understand why a task is assigned to a specific worker, and cannot trace whether the allocation result conforms to reasonable spatiotemporal constraints or user preferences. When a large number of tasks are overdue or workers frequently refuse, it is difficult to pinpoint the cause and correct the model. Fourth, generalization ability is limited. Many methods are designed for specific urban areas or specific task types, with highly customized model features and assumptions. When migrated to new cities or new application scenarios, their performance drops sharply, requiring data to be collected and retrained.

[0005] Therefore, there is an urgent need for a new task allocation framework that can organically integrate data-driven and knowledge-driven approaches. Summary of the Invention

[0006] To address the challenges of co-optimization between dynamic spatiotemporal constraints and multi-objective conflicts in task allocation within mobile crowdsourcing sensing networks, this invention provides a crowdsourcing sensing task allocation method and system based on decoupling of cue words and inference. By constructing a cue-driven task scheduling and inference framework integrating symbolic computation and neural reasoning, efficient screening and multi-objective optimized allocation of large-scale worker-task sets are achieved.

[0007] In a first aspect, the present invention provides a crowd-sensing task allocation method based on the decoupling of prompt words and inference, which adopts the following technical solution:

[0008] A crowd-sensing task allocation method based on decoupling of cue words and inference, comprising:

[0009] Obtain task data and worker data;

[0010] Multi-objective task modeling and candidate worker stratification are performed based on the acquired data;

[0011] Based on modeling and hierarchical results, local selection is performed within-group inference based on prompt words, including structured representation of candidate worker attributes, prompt-based multidimensional evaluation of individuals, intra-group cross-worker comparison and trade-off inference, and structured output and robust selection mechanism.

[0012] Global optimization under budget constraints is performed based on the reasoning results, including candidate set construction and task-level profile organization, multi-solution generation based on prompt words and mind tree reasoning, and selection of the global optimal solution under budget constraints.

[0013] Output the optimal solution.

[0014] Secondly, a crowd-sensing task allocation system based on the decoupling of prompt words and inference includes:

[0015] The data acquisition module is configured to acquire task data and worker data.

[0016] The stratification module is configured to perform multi-objective task modeling and candidate worker stratification based on the acquired data.

[0017] The selection module is configured to perform local selection based on prompt-word-based intra-group reasoning according to modeling and hierarchical results, including structured representation of candidate worker attributes, prompt-based multidimensional evaluation of individuals, intra-group cross-worker comparison and trade-off reasoning, and structured output and robust selection mechanism.

[0018] The optimization module is configured to perform global optimization under budget constraints based on the reasoning results, including candidate set construction and task-level profile organization, multi-solution generation based on prompt words and mind tree reasoning, and selection of the global optimal solution under budget constraints.

[0019] The output module is configured to output the optimal solution.

[0020] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned crowd intelligence perception task allocation method based on prompt words and decoupling.

[0021] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a crowd-sensing task allocation method based on prompt words and decoupling.

[0022] In summary, the present invention has the following beneficial technical effects:

[0023] 1. This invention significantly improves the efficiency and scalability of task allocation in large-scale dynamic scenarios by decoupling symbolic computation from neural reasoning. Specifically, before reasoning decisions, the feasibility of each worker-task pair is filtered and its potential is assessed. Based on the potential score, candidate workers are divided into high, medium, and low tiers, constructing a hierarchical candidate set that balances quality and diversity. This mechanism effectively avoids directly inputting massive amounts of low-potential or infeasible matches into a large language model for reasoning, thereby significantly reducing the search space and computational overhead of subsequent reasoning modules. This allows the invention to maintain efficient and stable scheduling performance even when facing large-scale worker-task sets and dynamic spatiotemporal constraints.

[0024] 2. This invention achieves effective collaborative optimization of multi-objective conflicts, overcoming the technical shortcomings of traditional methods that struggle to simultaneously consider conflicting indicators such as spatiotemporal feasibility, service quality, worker willingness, and budget constraints. In the local selection phase, it comprehensively weighs service quality, preference matching, and timeliness scores through structured representation of candidate worker attributes, prompt-based multidimensional individual evaluation, and intra-group cross-worker comparative reasoning. In the global optimization phase, it combines budget constraints and mind tree reasoning to coordinate multiple locally optimal solutions across tasks. Therefore, this invention can maximize task completion quality and worker willingness to perform, obtaining a globally optimal allocation scheme while satisfying spatiotemporal feasibility and budget constraints.

[0025] 3. This invention significantly enhances the interpretability and scenario adaptability of the task allocation process in crowd-sensing. Compared to traditional methods that rely on complex mathematical formulas, meticulously designed reward functions, or scenario-specific rules, this invention employs a prompt-based intra-group reasoning and multi-solution generation mechanism. This enables the large language model to perform explicit comparisons, trade-offs, and decision-making reasoning based on structured input and output. The allocation criteria are presented in natural language, facilitating understanding and intervention by maintenance personnel. Furthermore, by flexibly adjusting the prompt templates and weighting coefficients in the potential scoring parameters, this invention can adapt to various application scenarios, including different task types, worker preference distributions, urgency levels, and budget constraints, demonstrating good versatility and transferability. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a crowd-sensing task allocation method based on the decoupling of prompt words and inference in Embodiment 1 of the present invention.

[0027] Figure 2 This is a flowchart of a crowd-sensing task allocation method based on the decoupling of prompt words and inference in Embodiment 1 of the present invention.

[0028] Figure 3 This is a schematic diagram of individual multidimensional assessment based on prompt words in Embodiment 1 of the present invention.

[0029] Figure 4 This is a schematic diagram of the global optimization module under budget constraints in Embodiment 1 of the present invention.

[0030] Figure 5 This is a service quality performance graph of City A dataset under budget constraints in Embodiment 1 of the present invention.

[0031] Figure 6 This is a service quality performance graph of the city B dataset under budget constraints in Embodiment 1 of the present invention.

[0032] Figure 7 This is a service quality performance graph of the city C dataset under budget constraints in Embodiment 1 of the present invention.

[0033] Figure 8 This is a graph showing the execution intention performance of City A dataset under budget constraints in Embodiment 1 of the present invention.

[0034] Figure 9 This is a graph showing the execution intention performance of the city B dataset under budget constraints in Embodiment 1 of the present invention.

[0035] Figure 10 This is a graph showing the execution intention performance of the city C dataset under budget constraints in Embodiment 1 of the present invention.

[0036] Figure 11 This is a hypervolume representation of the city A dataset under budget constraints in Embodiment 1 of the present invention.

[0037] Figure 12 This is a hypervolume representation of the city B dataset under budget constraints in Embodiment 1 of the present invention.

[0038] Figure 13 This is a hypervolume representation of the city C dataset under budget constraints in Embodiment 1 of the present invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] Example 1

[0041] Reference Figure 1 This embodiment of a crowd-based intelligent perception task allocation method based on decoupling of prompt words and inference includes: "a crowd-based intelligent perception task allocation method based on decoupling of large model prompt words and computational inference", the overall technical framework of which includes four highly coupled core modules: (1) multi-objective task modeling and candidate worker hierarchical module (2) prompt word-based intra-group inference local selection module (3) budget constraint global optimization module, the specific scheme is as follows:

[0042] (1) Multi-objective task modeling and candidate worker hierarchical module

[0043] This module ( Figure 1 This model serves as the foundation for the entire mobile crowdsourcing perception task allocation system, serving as the modeling and decision-making basis. Its core task is to provide a unified formal description of the crowdsourcing workers participating in the allocation, the perception tasks, and their matching relationships. Based on this, a multi-objective optimization model is constructed that considers completion efficiency, preference consistency, service quality, and timeliness. Specifically, the system first abstracts the individual attributes of workers into a structured representation including elements such as location, online time, and movement speed. Simultaneously, tasks are characterized as standardized objects containing constraints such as task type, spatial location, and execution time window. Furthermore, key evaluation indicators are defined to quantitatively characterize the worker-task matching relationship from multiple dimensions, including spatial accessibility and temporal feasibility. To reduce the search complexity in the subsequent inference stage, the system first performs feasibility filtering on all worker-task pairs during the calculation stage, eliminating invalid matches that do not meet spatial, temporal, and availability conditions. Then, it integrates service quality, preference matching degree, and timeliness scores to construct a unified potential value, ranks and stratifies candidate workers, and finally generates a compact and high-quality candidate set for each task, thus providing reliable support for subsequent refined allocation decisions based on a large model.

[0044] 1) Definition of Crowdsourced Workers

[0045] To provide a unified model for the participants in the mobile crowdsourcing sensing task allocation process, we first formally define crowdsourcing workers. Crowdsourcing workers refer to individual users or device entities that actively or passively perform sensing tasks using smart terminal devices in a mobile crowdsourcing sensing environment. Let the set of crowdsourcing workers in the system be represented as:

[0046]

[0047] Any crowdsourced worker wi can be characterized as the following attribute tuple:

[0048]

[0049] in, A unique identifier for crowdsourced workers, used to distinguish different participating individuals; Indicates the type of task preferred by workers. This indicates the strength of their preference for this type of task. This attribute can reflect workers' subjective inclination in task selection and plays an important role in improving task matching satisfaction and willingness to perform. This indicates the worker's initial position, used to characterize their spatial starting point, and serves as the basis for subsequent accessibility analysis and path cost calculation; This indicates the online serviceable time range of workers. This time window reflects the availability constraints of workers within the task allocation cycle and is an important basis for judging whether a task can be executed on time.

[0050] Furthermore, and These represent the worker's reputation score and data quality score, respectively. The reputation score primarily reflects the worker's reliability in fulfilling historical tasks and service stability, and can be used to measure the credibility of their task completion. The data quality score characterizes the accuracy, completeness, and validity of the sensed data they upload, and is crucial for ensuring the overall usability of the sensed results. This indicates the worker's moving speed, used to estimate the time required for them to reach the task location; This represents the unit travel cost, used to quantify the movement expenses incurred by workers during task execution. This indicates the maximum acceptable service radius for a worker, used to constrain their task acceptance range and reflect the worker's service boundaries and task acceptance capacity in space.

[0051] Through the unified description of the aforementioned multi-dimensional attributes, crowdsourcing workers are not only modeled as execution subjects with basic identity information, but also represented as comprehensive decision-making units possessing characteristics of spatial accessibility, temporal feasibility, preference matching, service reliability, and execution cost. This formal definition lays the foundation for subsequent construction of worker-task matching relationships, design of multi-objective optimization functions, and implementation of high-quality task allocation strategies.

[0052] 2) Task Definition

[0053] In mobile crowdsourcing sensing systems, a task is defined as an atomic sensing activity with explicit spatial and temporal constraints, typically completed independently by a single worker. Let the set of all tasks in the system be denoted as . For any task It can be formally represented as an octet:

[0054] ,

[0055] in, A unique identifier for the task, used to distinguish different perception requests; Indicates the task type, which describes the perception category to which the task belongs, such as environmental monitoring, image acquisition, or location reporting; This indicates the target execution location of the task, reflecting its spatial deployment requirements. Tasks can only be executed within a time window. The internal execution, among which and These correspond to the start and end times of the task, respectively, and this constraint determines the timeframe for task allocation and execution. Parameters This indicates the reward offered by the task publisher for the task, used to incentivize workers to participate; This indicates the spatial coverage radius of a task; that is, as long as the worker is within this range, the spatial execution conditions of the task can be considered met. This indicates the urgency of the task. The higher the value, the higher the timeliness requirement for the task and the higher the priority it should be given in the task scheduling and allocation process.

[0056] The aforementioned parameters, from three levels—identification attributes, spatiotemporal attributes, and incentive attributes—combined to characterize the complete features of the task, enabling the system to model and manage tasks within a unified mathematical framework. In particular, the time window and urgency coefficient reflect the task's time sensitivity, which is crucial for identifying urgent perception requests requiring priority response; while location and coverage radius precisely describe the spatial accessibility of the task, providing a basis for subsequent worker-task matching. Through this structured task definition approach, the system can abstract complex and heterogeneous perception needs in real-world scenarios into standardized objects with clear physical meaning and optimization constraints, thus laying a solid foundation for subsequent task allocation, path planning, and incentive mechanism design.

[0057] 3) Candidate stratification based on competence and potential assessment

[0058] To explicitly improve the efficiency and matching quality of candidate screening in the task allocation phase, the computation phase and the reasoning phase are first decoupled before the reasoning decision-making phase. In the computation phase, for each worker-task pair... The process involves two complementary steps: feasibility filtering and capability potential assessment. Feasibility filtering eliminates allocation combinations that do not meet hard constraints, including spatial accessibility, time compatibility, overlapping available time periods, and duplicate allocation avoidance. For all candidate workers that pass the filtering, their corresponding uniform potential scores are further calculated to quantify the workers' capabilities. Execute the task The overall degree of adaptation can be expressed as:

[0059] ,

[0060] in, Depicting workers In the mission The service quality level on the platform comprehensively reflects factors such as its reputation, data quality, and willingness to perform, which is crucial for identifying highly reliable and stable candidate executors. Characterizing the degree of match between task type and worker interests or expertise helps to identify candidates with greater advantages in task semantics and ability structure; while Then assess the workers Whether a worker can arrive and complete the task within the effective time window reflects its feasibility under time constraints. By weighted fusion of the three key factors mentioned above, the potential score can uniformly model candidate workers from three dimensions: reliability, preference consistency, and time limit satisfaction. Subsequently, all feasible candidates are sorted according to their potential scores. When the candidate size is small, the entire candidate set is directly retained for subsequent inference. When the candidate size is large, a hierarchical partitioning strategy is further adopted, dividing them into three levels: high potential, medium potential, and low potential. Representative candidates are extracted from each level to form a hierarchical candidate set that balances quality and diversity, serving as input for the subsequent inference module. This preprocessing method allows the model to make decisions and inferences directly on a candidate set with clear semantic hierarchy after constraint filtering and potential evaluation, without having to perform an unconstrained search in a large-scale original candidate space. This significantly improves the efficiency, robustness, and interpretability of subsequent task allocation.

[0061] (2) Local selection module for in-group reasoning based on prompt words

[0062] This module ( Figure 2 This module is the core of the refined decision-making process in the entire task allocation system. Its main objective is to select the local set of workers that best meets the task requirements within each candidate worker group, considering multiple conflicting optimization objectives. Unlike traditional methods that rely on fixed weighting, linear scoring functions, or heuristic ranking rules, this module introduces a prompt-based intra-group reasoning mechanism. This explicitly injects multi-objective decision-making logic into the reasoning process of the large language model, enabling the model to perform fine-grained analysis and comparison of the comprehensive capabilities of candidate workers in complex contexts. Specifically, the system no longer treats factors such as quality, timeliness, willingness to execute, benefit level, and preference matching degree independently. Instead, it guides the model to jointly evaluate the multi-dimensional attributes of each candidate worker from a holistic perspective through unified structured prompts, thereby generating more reasonable, adaptive, and interpretable ranking results.

[0063] 1) Structured Representation of Candidate Worker Attributes: To overcome the problem of discrete and heterogeneous original candidate worker information and the difficulty in direct comparison, this invention proposes a structured attribute representation method for candidate workers based on potential energy hierarchies. For each task... First, construct its corresponding hierarchical candidate set. The specific construction process is as follows.

[0064] For each worker-task pair The four hard constraints are checked sequentially. Spatial accessibility requirements for workers... The service radius can cover the task location:

[0065]

[0066] in For Euclidean or road network distances. The maximum service radius for workers. Time compatibility requires that workers expect to complete tasks no later than the task deadline.

[0067]

[0068] The availability overlap constraint ensures that there is an overlap between the worker's idle time and the effective time window of the task:

[0069]

[0070] The redundancy avoidance constraint ensures that workers have not been assigned to other unfinished tasks. If any constraint is not satisfied, the pair is directly discarded and not included in subsequent calculations. For feasible pairs that pass the filtering, a uniform potential energy fraction is calculated. Used to quantify workers For the task Overall compatibility:

[0071]

[0072] Where the weight coefficients satisfy In one embodiment of the present invention, This emphasizes the priority of ensuring the reliability of task completion. The potential energy score is composed of three weighted components: service quality component... A comprehensive reflection of workers' historical reliability and willingness to perform, determined by a credit score. Data quality And willingness coefficient Weighted average yields:

[0073]

[0074] in Willingness coefficient It can be dynamically estimated based on the frequency with which workers have historically received similar tasks. Preference matching component. To measure the alignment between task type and worker interests / expertise, task design is performed. The type vector is The worker's preference vector is ,but:

[0075]

[0076] Time urgency component The feasibility and sufficiency of evaluating whether workers can complete a task within the effective time window are defined as:

[0077]

[0078] in As a time relaxation factor, for each task , calculate Sort in descending order to obtain an ordered list. .like ( (If the minimum stratification threshold is used), then directly let Otherwise, Based on the potential energy fraction, it is divided into three segments: high, medium, and low. For Each of the candidate workers The system extracts its relationship with the task. The relevant key attributes are organized into the following multi-dimensional vector:

[0079]

[0080] This profile comprehensively depicts the worker's potential adaptability under multi-objective constraints, encompassing dimensions such as task completion quality, response timeliness, execution willingness, individual benefits, and the degree of matching with task preferences. Compared to directly using raw numerical values ​​or simple ratings, the structured representation described above transforms discrete and heterogeneous raw information into semantically clear and dimensionally unified descriptive units. Furthermore, through potential scores and hierarchical sampling mechanisms, it preserves the diversity and representativeness within the candidate set, providing standardized, high-information-density input for subsequent reasoning analysis based on large language models.

[0081] 2) Prompt-based multidimensional individual assessment

[0082] After completing the candidate worker attribute modeling, the system performs each task Construct a structured candidate set ,in For the task The number of candidate workers. Each worker The attribute vector is represented as:

[0083]

[0084] These correspond to quality, timeliness, willingness, benefits, and the task. The degree of preference matching. Let the global optimization objective set be... Each of the targets It can be formalized as a partial order relation or utility function. The system will identify the task. Candidate worker attribute set and target set Commonly inject prompt templates to generate task-oriented... Reasoning hints The construction process of the prompt follows a structured template, which includes a task description, a list of candidate workers (affiliation vector), a description of the optimization objective, and explicit inference instructions.

[0085] Reasoning hints The model is guided to perform the first two stages of the four-stage reasoning process: dimensional labeling and individual comprehensive evaluation. The model first qualitatively labels each attribute of each candidate worker. The label mapping function is defined as follows:

[0086]

[0087] Each of them For a finite set of discrete labels, such as Corresponding to three quality levels: high, medium, and low; Corresponding to fast, medium, and slow delivery times; Corresponding to strong, medium, and weak intentions; Corresponding to high, medium, and low returns; This corresponds to high, medium, and low preferences. This mapping is achieved through a preset quantile threshold or rule function; for example, for the quality dimension:

[0088]

[0089] in and For the preset threshold, satisfy The same applies to other dimensions. Through this discretized representation, the model obtains an interpretable coarse-grained worker profile, providing a semantic basis for subsequent comprehensive evaluation.

[0090] 3) Intragroup cross-worker comparison and trade-off reasoning

[0091] After completing individual assessments, the system further guides the large language model to conduct cross-worker comparative analysis within the candidate worker set. Unlike traditional methods that rank workers solely based on a single total score, this stage emphasizes joint reasoning regarding the balance, complementarity, and dominant strengths of candidate workers across multiple objective dimensions. The model not only needs to identify a worker's outstanding performance on a single indicator but also needs to comprehensively analyze whether they possess superior multidimensional coordination capabilities under the overall task requirements. This comparative reasoning mechanism plays a crucial role, especially in scenarios where multiple objectives conflict. For example, when there is an inherent contradiction between high quality and low cost, or between rapid response and high preference matching, the model can explicitly analyze the trade-offs between different candidates by incorporating the task context, thereby determining whether to select a worker who excels in one indicator but is unbalanced overall, or a worker with a more balanced performance across multiple indicators and superior overall performance. This reasoning process ensures that the final selection result no longer relies on static weight settings but is more closely aligned with the actual needs of the task itself.

[0092] Therefore, the system will prompt... The document explicitly requires the model to conduct cross-worker comparative analysis for all candidate workers. Define the candidate worker set. The model is tailored to each worker. Individual assessment vectors have been obtained. Building on this, the model further compares any two workers. and The comparison process assesses the relative performance across each objective dimension and outputs a comparative conclusion. This comparison can be formally represented as:

[0093]

[0094] in To optimize the target set, The model uses natural language reasoning to explicitly handle scenarios where multiple objectives conflict. For example, when there is an inherent contradiction between high quality and low cost, or between rapid response and high preference matching, the model can combine the task context to explicitly analyze the trade-offs between different candidates, thereby determining whether to choose a worker who excels in one indicator but is unbalanced overall, or a worker who is more balanced across multiple indicators and has a better overall performance.

[0095] After completing all pairwise comparisons, the model generates a total or partial order relation for the candidate workers. The final output is an ordered top-order relation that meets the task requirements. List , such that for any have ,in This represents the preference relationship derived from the cross-worker comparative reasoning described above. Through this comparative reasoning, the final selection result no longer relies on static weight settings, but rather better reflects the actual needs of the task itself.

[0096] 4) Structured output and robust selection mechanism

[0097] After completing the intra-group comprehensive comparison, the large language model outputs the final decision result in a unified structured format. (Task description follows) The final selection result is The output not only includes the ranking of candidate workers but can also include a brief summary of the selection criteria and attributes, thereby improving the interpretability and traceability of the screening results. The output can be formally represented as:

[0098]

[0099] To ensure system stability and engineering usability, when the model returns a valid result containing a complete sort, the system directly parses its response. Generate the final allocation set Define the legality conditions:

[0100]

[0101] when At that time, the system parsed and obtained When the output is missing, has an abnormal format, or is inconsistent with the task objective, a preset fallback rule is triggered, and supplementary filtering is performed from the candidate attribute set in a deterministic manner. This mechanism effectively compensates for the shortcomings of language models in terms of output standardization and stability, ensuring that the entire local selection module still has good robustness and reliability in complex task scenarios.

[0102] (3) Global optimization module under budget constraints

[0103] This module ( Figure 3 , Figure 4 This is the core decision-making module that coordinates and optimizes the global task allocation results after the prompt-based local reasoning within the group is completed. Its main objective is to jointly select the final executor from the local candidate sets corresponding to all tasks, based on the high-quality candidate worker sets already obtained for each task, forming a global task-worker allocation scheme that satisfies a unified incentive budget constraint. Unlike the previous stage which evaluated each task independently, this module is no longer limited to a single-task perspective, but rather considers multiple dimensions such as quality, timeliness, cost, willingness, and preference matching from a global perspective across tasks and candidate sets, within the budget limit. To maximize overall allocation utility under constraints. Formally, let the task set be... The candidate worker group is The core problem that needs to be solved in this stage is: for each task Select a single worker from its candidate set and construct the final allocation scheme. This ensures that the total incentive cost does not exceed the budget. At the same time, the overall allocation quality is optimal. Unlike traditional combinatorial optimization methods that rely on integer programming, branch and bound, or heuristic search, this module models this global coordination process as a tree-like decision-making process driven by a large language model. This enables the model to explicitly weigh the conflicts and synergies between multiple objectives in natural language reasoning, thereby outputting an optimal allocation scheme that is feasible, globally consistent, and interpretable.

[0104] 1) Candidate set construction and task-level profiling organization

[0105] Before entering the global optimization phase, the system first performs optimization for each task based on the intra-group inference results from the previous phase. Construct the corresponding candidate worker set Furthermore, in order for the language model to understand the capabilities and suitability of each candidate worker, it is necessary to organize the multidimensional attributes of each candidate worker into a structured profile. Specifically, for each candidate worker... Construct a 1D attribute vector

[0106]

[0107] in Indicates historical quality level, Indicates the expected response time. Indicates the required incentive cost. Indicates subjective willingness to participate. This indicates the degree of matching with task preferences. All profile vectors are organized into a three-dimensional tensor.

[0108]

[0109] Positions where the candidate set is insufficient are padded with zero vectors. This portrait tensor serves as the fundamental input for subsequent cue construction, transforming discrete candidate workers into semantic decision units that can be understood and compared by the inference model.

[0110] 2) Multi-solution generation and mind tree reasoning based on prompt words

[0111] After obtaining task-level candidate profiles, the system constructs prompts for each task. This prompt is generated by the image tensor. Task Description and the optimization target set Together, that is

[0112]

[0113] By explicitly injecting task context, candidate worker attributes, and global optimization goals (including budget cap) (And worker mutual exclusion constraints), guiding the large language model to generate multiple possible task allocation schemes under a unified constraint framework. The model receives prompts. Then, output the response. And from this, a set of candidate allocation schemes is extracted.

[0114]

[0115] Each Each solution represents a feasible task-worker matching scheme, sharing the fundamental constraint that "each task must be assigned to one and only one worker." Different schemes reflect different optimization tendencies; for example, some schemes emphasize the lowest overall cost, some emphasize higher completion quality and task reliability, while others focus more on worker load balancing or user preference consistency. This multi-solution generation mechanism is essentially equivalent to a prompt-guided tree expansion in the solution space: each candidate scheme corresponds to a different global decision path, and the large language model explores these paths through semantic reasoning capabilities. To quantify the differences between schemes, a branch divergence degree can be defined.

[0116]

[0117] For each generated candidate solution The system further calculates its multidimensional utility vector.

[0118]

[0119] The components are defined as follows:

[0120]

[0121] The overall score is defined as a weighted sum:

[0122]

[0123] Where the weight vector satisfy It can be dynamically configured according to the task scenario. During the mind tree reasoning process, only those scores above a threshold are considered valid. Only the best solutions will proceed to the next round of evaluation, thus enabling heuristic pruning.

[0124] 3) Selection of the globally optimal solution under budget constraints

[0125] After interpreting and evaluating all candidate solutions, the system enters the final global selection phase. First, it collects all candidate solutions generated by the tasks to obtain a global candidate solution set. Then, solutions that violate global budget constraints or worker conflict constraints are removed to obtain a set of feasible solutions.

[0126] Next, the system will list each feasible solution Together with its utility vector and natural language interpretation generated by the model Input into the final evaluation prompt. The large language model is required to perform pairwise comparisons:

[0127]

[0128] Furthermore, the preference relationships are output, and a preference graph is constructed by collecting all pairwise comparison results. ,in express Superior Then extract the Pareto front.

[0129]

[0130] If the Pareto set contains multiple solutions, then the large language model further selects a final solution based on a comprehensive trade-off:

[0131]

[0132] Final output The feasibility requirement must be met simultaneously (total incentive expenditure does not exceed the preset upper limit). The optimality requirement (to improve the overall quality of task allocation as much as possible within the budget and achieve a balance and coordination among multi-dimensional goals) is also considered. From a decision-making logic perspective, this stage does not simply select the option with the highest single indicator, but rather uses a language model to comprehensively compare the multi-dimensional scores, explanatory information, and goal trade-offs of each candidate option, ultimately selecting the most balanced and globally consistent task allocation result.

[0133] Experimental verification

[0134] To comprehensively evaluate the performance of the proposed method, this paper selects six representative mainstream methods as comparative baselines. MaOWS is a multi-objective optimization framework based on augmented differential evolution, capable of jointly optimizing worker benefits, platform benefits, requester payments, and perceived quality in large-scale crowdsourcing systems; FFU is a greedy heuristic allocation method that iteratively balances worker utility and platform profit to complete task allocation; HHO is a swarm intelligence metaheuristic algorithm that combines global search and local exploitation capabilities; C-ILP models the worker-task allocation problem as a constrained integer linear programming problem to minimize incentive payments and movement distances; IO is a naive single-step method based on a large language model that inputs all workers, tasks, and optimization objectives into the model at once without separating reasoning and computation; Random is a random allocation strategy that does not consider any optimization objective and is used as a lower bound baseline for performance.

[0135] To evaluate the task allocation performance of various methods from different perspectives, this paper uses metrics such as service quality, execution willingness, hypervolume index, and token consumption for comprehensive evaluation. Service quality measures the quality of the matching results between workers and tasks; execution willingness reflects the spatial and temporal feasibility of task allocation and the likelihood of workers actually accepting tasks; hypervolume index measures the overall superiority of the Pareto solution set in the context of multi-objective optimization; and token consumption evaluates the inference cost and scalability of large language model-driven methods during the solution process.

[0136] In service quality indicators ( Figure 5 , Figure 6 , Figure 7 On the datasets of cities A, B, and C, the method of this invention consistently demonstrates superior performance, indicating its ability to reliably identify better worker-task matching relationships and achieve higher-quality task execution results. Especially in scenarios with low budget constraints, the negative impact of low-quality allocation decisions is amplified by limited resources, making the advantages of the method of this invention even more significant. In contrast, heuristic or optimization methods such as FFU, HHO, and C-ILP generally exhibit significant fluctuations as budgets increase. This is mainly due to their relatively rigid target aggregation methods or the potential for introducing biases in local selection processes, leading to unstable overall allocation quality. While the IO method possesses some semantic reasoning capabilities, it lacks an effective structured candidate screening and phased decision-making mechanism, making it difficult to maintain stable, high-quality output in complex task allocation scenarios. The Random method, lacking any optimization capabilities, performs the worst overall. In contrast, the method of this invention can jointly consider multi-dimensional information such as worker reputation, willingness to perform, and data quality, and through a "computation-reasoning decoupling" design, it makes more accurate decisions within the candidate space, thus maintaining high and stable performance under different budget conditions.

[0137] In terms of willingness to implement ( Figure 8 , Figure 9 , Figure 10On the three datasets of City A, City B, and City C, the method of this invention comprehensively outperforms all comparative methods. Execution willingness reflects the probability of workers completing tasks under various constraints such as spatial distance, time window, and task acceptance tendency, and is a key factor affecting the success rate of crowdsourcing tasks. As budgets increase, many comparative methods, in order to expand task coverage, often gradually allocate tasks to workers with weaker spatial or temporal matching, resulting in a decrease in execution willingness. Traditional heuristic and optimization methods, in particular, are more prone to imbalances between increased revenue and feasibility constraints. In contrast, the method of this invention can explicitly model worker execution willingness during the allocation process, incorporating distance, urgency, worker availability, and preference factors into a unified analytical framework, and achieving a joint trade-off between accessibility and acceptance probability through a prompt-based multidimensional reasoning mechanism. Therefore, even with increased budgets and more complex candidate allocation combinations, the method of this invention can still maintain a high and stable level of execution willingness, demonstrating its superior ability to balance task coverage and execution feasibility.

[0138] In the hypervolume index ( Figure 11 , Figure 12 , Figure 13 On all datasets, the method of this invention achieves the largest hypervolume value, and its advantage over the comparative methods further expands as the budget increases. Hypervolume is a typical indicator for measuring the overall performance of multi-objective optimization, reflecting both the convergence and distribution of the solution set across multiple objective dimensions. A larger hypervolume value means that the obtained solution achieves a better overall trade-off among multiple competing optimization objectives. Experimental results show that the method of this invention can not only find higher-quality local solutions, but also maintain effective exploration and approximation of the global Pareto front even when the budget increases and the solution space expands significantly. In contrast, evolutionary methods such as MaOWS and HHO, while possessing certain global search capabilities, are still prone to getting trapped in local trade-off regions under complex high-dimensional constraints; exact optimization methods such as C-ILP are limited by modeling rigidity and solution complexity, and their Pareto solution set expansion capability is limited when facing real-world scenarios with multiple attributes, preferences, and constraints; the IO method, due to its failure to achieve structured decomposition and constraint-driven screening, also struggles to consistently output high-quality solutions at the multi-objective level. Therefore, the superior performance of the method of this invention in terms of hypervolume index fully demonstrates that its global reasoning mechanism under budget constraints can more flexibly and comprehensively coordinate multiple optimization objectives, thereby achieving better overall allocation efficiency.

[0139] Table 1. Comparison of token consumption for different methods across various datasets.

[0140]

[0141] In terms of token consumption metrics (Table 1), the method of this invention significantly outperforms the naive IO baseline under the maximum budget settings of each dataset, while also completing more tasks, demonstrating outstanding cost advantages and scalability. Specifically, the IO method, by encoding all workers, tasks, and optimization objectives into a single prompt at once, experiences a rapid increase in prompt length as the problem size expands, leading to a large amount of redundant reasoning and a significant out-of-control trend in token consumption. In contrast, the method of this invention first prunes infeasible or low-quality worker-task combinations during the computation phase, and then performs structured reasoning within the reduced candidate space, effectively avoiding invalid analysis of irrelevant combinations, significantly compressing the prompt length, and keeping the token cost per task relatively stable. This indicates that the method of this invention not only outperforms existing methods in terms of allocation quality but also has stronger robustness, scalability, and cost-effectiveness in large-scale real-world crowdsourcing scenarios.

[0142] Example 2

[0143] This embodiment provides a crowd-aware task allocation system based on the decoupling of prompt words and inference, including:

[0144] The data acquisition module is configured to acquire task data and worker data.

[0145] The stratification module is configured to perform multi-objective task modeling and candidate worker stratification based on the acquired data.

[0146] The selection module is configured to perform local selection based on prompt-word-based intra-group reasoning according to modeling and hierarchical results, including structured representation of candidate worker attributes, prompt-based multidimensional evaluation of individuals, intra-group cross-worker comparison and trade-off reasoning, and structured output and robust selection mechanism.

[0147] The optimization module is configured to perform global optimization under budget constraints based on the reasoning results, including candidate set construction and task-level profile organization, multi-solution generation based on prompt words and mind tree reasoning, and selection of the global optimal solution under budget constraints.

[0148] The output module is configured to output the optimal solution.

[0149] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned crowd-sensing task allocation method based on decoupling of prompt words and inference.

[0150] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a crowd-sensing task allocation method based on prompt words and decoupling.

[0151] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for crowd-sensing task allocation based on coupling of cues and reasoning, characterized in that, include: Obtain task data and worker data; Multi-objective task modeling and candidate worker stratification are performed based on the acquired data; Based on modeling and hierarchical results, local selection is performed within-group inference based on prompt words, including structured representation of candidate worker attributes, prompt-based multidimensional evaluation of individuals, intra-group cross-worker comparison and trade-off inference, and structured output and robust selection mechanism. Global optimization under budget constraints is performed based on the reasoning results, including candidate set construction and task-level profile organization, multi-solution generation based on prompt words and mind tree reasoning, and selection of the global optimal solution under budget constraints. Output the optimal solution; The multi-objective task modeling and candidate worker stratification based on the acquired data also includes decoupling the computation phase from the inference phase before reasoning and decision-making, for each worker-task pair, in order to explicitly improve the efficiency and matching quality of candidate screening in the task allocation phase. The process involves two complementary steps: feasibility filtering and capability potential assessment. For all candidate workers who pass the filtering, a uniform potential score is calculated to quantify the worker's capabilities. Execute the task The overall degree of adaptation is expressed as: ， in, Depicting workers In the mission The level of service quality; Characterizes the degree of match between task type and worker interests or skills / expertise; worker assessment Whether the task can be completed in time within the effective time window; then all feasible candidates are sorted according to their potential scores, and a hierarchical partitioning strategy is adopted to divide them into three levels: high potential, medium potential and low potential. Representative candidates are extracted from each level to form a hierarchical candidate set that takes into account both quality and diversity, which serves as the input for the subsequent reasoning module. The multi-solution generation and mind tree reasoning based on prompt words include constructing prompts for each task after obtaining task-level candidate profiles. This prompt is generated by the image tensor. Task Description and the optimization target set Together they form, namely: By explicitly injecting task context, candidate worker attributes, and global optimization objectives, the large language model is guided to generate multiple possible task allocation schemes under a unified constraint framework, and the model receives prompts. Then, output the response. And from this, a set of candidate allocation schemes is extracted: each Each represents a feasible task-worker matching scheme. To quantify the differences between different schemes, the degree of divergence is defined: For each generated candidate solution its multidimensional utility vector is computed: where the components are defined as follows: , The composite score is defined as a weighted sum: where the weight vector satisfies ; The selection of the globally optimal solution under the budget constraint includes, after interpreting and evaluating all candidate solutions, collecting all candidate solutions generated by the task to obtain a global candidate solution set. Then, solutions that violate global budget constraints or worker conflict constraints are removed to obtain a set of feasible solutions: ， Each feasible solution is then input along with its utility vector and natural language explanations generated by the model into a final evaluation prompt, where the large language model is asked to make pairwise comparisons: Then, the preference relationships are output, and a preference graph is constructed by collecting all pairwise comparison results. ,in express Superior Then extract the Pareto front: ， If the Pareto set contains multiple solutions, the large language model selects the final solution based on a comprehensive trade-off: 。 2. The crowd-sensing task allocation method based on coupling prompt words and reasoning according to claim 1, wherein, The multi-target task modeling and candidate worker stratification based on the acquired data includes unified modeling for the participating subjects in the allocation process of the mobile crowd sensing task. First, the crowd workers are formally defined, and the crowd worker set is represented as: where any crowd worker wi is depicted as an attribute tuple as follows: ,in, A unique identifier for crowdsourced workers; Indicates the type of task preferred by workers. This indicates the strength of their preference for this type of task; Indicates the worker's initial position; This indicates the range of time during which workers are available to work online. and These represent the worker's credit score and the data quality score, respectively. Indicates the worker's moving speed; Indicates the cost per unit of travel; Let represent the maximum acceptable service radius for workers; then let the set of all tasks be denoted as . For any task Formalized as an octet: ,in, A unique identifier representing a task; Indicates the task type; Indicates the target execution location of the task; the task can only be executed within the time window. The internal execution, among which and These correspond to the start and end times of the task, respectively, and the parameters. This indicates the reward offered by the task poster for the task; Indicates the spatial coverage radius of the task; This indicates the urgency of the task.

3. The crowd-sensing task allocation method based on coupling prompt words and reasoning according to claim 2, characterized in that, The structured representation of candidate worker attributes includes a potential energy-based hierarchical structured attribute representation method to overcome the problems of discrete and heterogeneous original candidate worker information and the difficulty in direct comparison. This method is applied to each task. First, construct its corresponding hierarchical candidate set. The specific construction process is as follows: for each worker-task pair The four hard constraints are examined in sequence, including the spatial accessibility requirements for workers. The service radius can cover the task location: in For Euclidean or road network distances. To maximize the worker's service radius, time compatibility requires that the worker's expected completion time for the task is no later than the task's deadline. The availability overlap constraint ensures that there is an overlap between the worker's idle time and the effective time window of the task: The redundancy avoidance constraint ensures that workers have not been assigned to other unfinished tasks. If any constraint is not satisfied, the pair is directly eliminated. For feasible pairs that pass the filtering, a uniform potential energy fraction is calculated. Used to quantify workers For the task Overall compatibility: Among them, service quality component A comprehensive reflection of workers' historical reliability and willingness to perform, determined by a credit score. Data quality and willingness coefficient Weighted average yields: in Willingness coefficient Based on a dynamic estimation of the frequency with which workers have historically accepted similar tasks, the preference matching component is determined. To measure the alignment between task type and worker interests and expertise, tasks are set up. The type vector is The worker's preference vector is ,but: Time urgency component The feasibility and sufficiency of workers completing tasks within the effective time window are defined as follows: in As a time relaxation factor, for each task , calculate Sort in descending order to obtain an ordered list. ,like , To set the minimum stratification threshold, we can directly set... Otherwise, Based on the potential energy fraction, it is divided into three segments: high, medium, and low. each of the candidate workers extracts its key attributes relevant to the task organized as a multi-dimensional vector: 。 4. The crowd-sensing task allocation method based on coupling prompt words and reasoning according to claim 3, characterized in that, The prompt-based individual multidimensional assessment includes, after completing candidate worker attribute modeling, performing each task... Constructing a structured candidate set ,in For the task The number of candidate workers, each worker The attribute vector is represented as: These correspond to quality, timeliness, willingness, benefits, and relevance to the task, respectively. The preference matching degree, let the global optimization objective set be Each of the targets Formalize as a partial order relation or utility function , assign task identifier Candidate worker attribute set and target set Commonly inject prompt templates to generate task-oriented... Reasoning hints Reasoning hints To guide the large language model in executing the first two stages of the four-stage inference process—dimensional labeling and individual comprehensive evaluation—we first qualitatively label each dimension of each candidate worker's attribute, defining a label mapping function: Each of them For a finite set of discrete labels, Corresponding to fast, medium, and slow delivery times; Corresponding to strong, medium, and weak intentions; Corresponding to high, medium, and low returns; The mapping corresponds to high, medium, and low preferences, achieved through a preset quantile threshold. For the quality dimension: ， wherein with is a preset threshold value, satisfying .

5. The crowd-sensing task allocation method based on coupling prompt words and reasoning according to claim 4, characterized in that, The intragroup cross-worker comparison and trade-off reasoning and structured output and robust selection mechanism, including prompts after individual assessment, are provided. The process involves cross-worker comparative analysis of all candidate workers, defining the candidate worker set. For each worker Individual assessment vectors have been obtained. Compare any two workers and The relative performance across each objective dimension is expressed as follows: ,in To optimize the target set, Using the task context, after completing all comparisons, generate a total order relation on the candidate workers, and finally output an ordered top-order relation that satisfies the task requirements. List , such that for any have ,in This represents the preference relationship obtained based on the above cross-worker comparison reasoning; After completing the comprehensive comparison within the group, the final decision result will be output in a unified structured format. (Task description follows) The final selection result is The output is formally represented as: When the returned result is legal and contains a complete ordering, directly parse its response Generate the final allocation set Define the legality condition: ， When , the analysis obtains ; and when the output is missing, abnormal in format or inconsistent with the task target, the preset fallback rule is triggered to complete the supplementary screening in a deterministic manner from the candidate attribute set.

6. The crowd-sensing task allocation method based on coupling prompt words and reasoning according to claim 5, characterized in that, The candidate set construction and task-level profile organization include, before entering the global optimization stage, based on the intra-group inference results of the previous stage, creating a profile for each task. Construct the corresponding candidate worker set To enable the language model to understand the capabilities and suitability of each candidate worker, the multidimensional attributes of each candidate worker are organized into a structured profile. Specifically, for each candidate worker... Construct a 3D attribute vector: in Indicates historical quality level, Indicates the expected response time. Indicates the required incentive cost. Indicates subjective willingness to participate. Representing the degree of matching with task preferences, all profile vectors are organized into a three-dimensional tensor: Pad with zero vectors for positions with insufficient candidate set size.

7. A crowd-sensing task allocation system based on prompt-word and reasoning coupling, which executes a crowd-sensing task allocation method based on prompt-word and reasoning coupling according to claim 1. include: The data acquisition module is configured to acquire task data and worker data. The stratification module is configured to perform multi-objective task modeling and candidate worker stratification based on the acquired data. The selection module is configured to perform local selection based on prompt-word-based intra-group reasoning according to modeling and hierarchical results, including structured representation of candidate worker attributes, prompt-based multidimensional evaluation of individuals, intra-group cross-worker comparison and trade-off reasoning, and structured output and robust selection mechanism. The optimization module is configured to perform global optimization under budget constraints based on the reasoning results, including candidate set construction and task-level profile organization, multi-solution generation based on prompt words and mind tree reasoning, and selection of the global optimal solution under budget constraints. The output module is configured to output the optimal solution.

Images