Channel-aware occupational exposure assessment method, storage medium, and electronic device

Abstract

Embodiments of the present disclosure disclose a channel-aware occupational exposure assessment method, a storage medium and an electronic device. The method comprises: obtaining a plurality of task-level natural language description texts of a target occupation; for each task-level natural language description text, obtaining a model output result by using a plurality of large language model integration; constructing two independent capability frontier vectors, respectively representing the current capability level of software artificial intelligence and embodied robot in different cognitive dimensions and sensory-motor dimensions; calculating the basic exposure score of the target occupation based on the capability demand vector of the target occupation and the two capability frontier vectors; and performing channel-by-channel discount adjustment on the basic exposure score based on the channel weight distribution and the coupling strength to obtain an adjusted exposure score.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, specifically to a channel-aware occupational exposure assessment method, storage medium, and electronic device. Background Technology

[0002] With the rapid development of artificial intelligence and robotics, assessing the risk of various occupations being replaced by automation is crucial for workforce planning, education investment decisions, and individual career development. Existing technologies mainly fall into three categories: The first category is based on expert judgments of the "computerizability" of occupational tasks, combined with machine learning classifiers, to output a single synthetic exposure score for each occupation. This type of index has been widely adopted since 2017 and has become a foundational reference for automation exposure research. The second category maps advancements in AI capability benchmarks to occupational skill lists (such as the ESCO system), outputting a single AI exposure score for each occupation. Since 2021, this has become the mainstream reference for AI exposure research. The third category is an assessment method based on large language models, scoring occupational tasks item by item and then summing the scores. Using large language models as scorers, it generates multi-dimensional capability scores from occupational text descriptions. The common assumption of these three methods is that the capabilities required for an occupation can be decomposed into mutually independent capability dimensions. Each dimension is compared with the corresponding technological frontier level, and then linearly summed to obtain the total occupational exposure score. For ease of description, this is referred to here as the "additive separability assumption." Summary of the Invention

[0003] In view of this, embodiments of the present disclosure provide a channel-aware occupational exposure assessment method, storage medium, and electronic device.

[0004] In a first aspect, this disclosure provides a channel-aware occupational exposure assessment method, the method comprising the following steps: Obtain multiple task-level natural language description texts for the target profession; For each task-level natural language description text, multiple large language models are integrated to obtain model output results. The model output results include: coupling strength, channel weight distribution, and a pair of bound dominant cognitive dimension and dominant sensorimotor dimension. The coupling strength is used to characterize the degree to which the cognitive judgment of the corresponding task is formed instantly through real-time sensorimotor feedback. The channel weight distribution includes the channel weights of the corresponding task on multiple predefined perceptual-motor channels. The bound dominant cognitive dimension and dominant sensorimotor dimension are the predetermined cognitive dimension and predetermined sensorimotor dimension, respectively, when the cognitive judgment of the corresponding task can be formed instantly through real-time sensorimotor feedback. Two independent capability frontier vectors are constructed to represent the current capability levels of software artificial intelligence and embodied robots in different cognitive and sensorimotor dimensions, respectively. The basic exposure score of the target occupation is calculated based on the capability requirement vector of the target occupation and the two capability frontier vectors; the capability requirement vector is used to characterize the requirement level of the target occupation in different predetermined cognitive and sensorimotor dimensions; the basic exposure score is used to characterize the coverage of the capability level of the software artificial intelligence and the embodied robot in meeting the requirement level of the target occupation. The basic exposure score is adjusted by discounting each channel based on the channel weight distribution and the coupling strength to obtain the adjusted exposure score.

[0005] Furthermore, the method also includes: The coupling structure of the target occupation is obtained based on the coupling strength corresponding to the task and the channel weight distribution. The coupling structure includes the total coupling strength obtained by weighting the coupling strength of multiple tasks, the channel spectrum obtained by weighting the channel weight distribution of multiple tasks respectively, and the decoupling margin. The decoupling margin represents the proportion of tasks with a coupling strength of 0 for the target occupation.

[0006] Furthermore, the method also includes: Using the channel weight distribution as a sample, cluster the predefined multiple perception-action channels, and based on the clustering results, remove a set number of perception-action channels and their corresponding channel weights that have the fewest clusters from the channel weight distribution.

[0007] Furthermore, the method also includes: Constructing the future capability frontier vectors of software artificial intelligence and embodied robots at multiple points in the future; Based on the future capability frontier vector, the adjusted exposure score of the target profession is obtained at multiple future time points.

[0008] Furthermore, the method also includes: Obtain multidimensional competency assessment information for the target occupation from a designated database; The multidimensional ability assessment information is projected onto multiple predetermined cognitive dimensions and multiple predetermined sensorimotor dimensions using a predetermined mapping matrix to obtain the ability requirement vector for the target occupation.

[0009] Further, the predetermined mapping matrix includes a cognitive ability mapping sub-matrix and a sensorimotor ability mapping sub-matrix; the multidimensional ability assessment information is projected onto multiple predetermined cognitive dimensions and multiple predetermined sensorimotor dimensions using the predetermined mapping matrix to obtain the ability requirement vector for the target occupation, including: The cognitive ability assessment information in the multidimensional ability assessment information is projected onto multiple predetermined cognitive dimensions using the cognitive ability mapping sub-matrix to obtain cognitive sub-components; The cognitive sub-component is weighted and fused with the model output sub-component to obtain the cognitive demand component; the model output sub-component is obtained by evaluating the task-level natural language description text of the target occupation on the predetermined cognitive dimension through multiple large language models. The sensorimotor ability assessment information in the multidimensional ability assessment information is projected onto the predetermined sensorimotor dimension using the sensorimotor ability mapping submatrix to obtain the sensorimotor demand component. The cognitive demand component and the sensorimotor demand component are concatenated to obtain the capability demand vector.

[0010] Furthermore, the method also includes: Calculate the consistency among the outputs of the multiple large language models; the consistency includes the consistency of channel weight distribution, the consistency of coupling strength, and / or the consistency between the bound dominant cognitive dimension and the dominant sensorimotor dimension.

[0011] Furthermore, the method also includes: For each task-level natural language description text, the multiple large language models are used to further obtain the interest content assessment information of the target occupation in multiple predefined interest dimensions; The adjusted exposure score is adjusted using the benefit / harm content assessment information to obtain the benefit / harm corrected exposure score.

[0012] Further, the adjusted exposure score is adjusted using the aforementioned benefit content assessment information to obtain a benefit-corrected exposure score, including: Obtain the preset benefit / discount factor; The adjusted exposure score is corrected using the interest discount factor to obtain the interest-corrected exposure score.

[0013] Furthermore, the dominant cognitive dimension includes at least one of the following: verbal, fluid reasoning, quantity, spatial, working memory, speed attention, and creativity; the dominant sensorimotor dimension includes at least one of the following: fine manipulation, control coordination, reaction speed, near vision, far vision, color vision, hearing, strength, endurance, flexibility, and gross coordination.

[0014] In a second aspect, embodiments of this disclosure provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the above aspects.

[0015] Thirdly, embodiments of this disclosure provide an electronic device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described in any of the above aspects.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0017] The above and other objects, features, and advantages of this disclosure will become clearer from the following description of embodiments of this disclosure with reference to the accompanying drawings, in which: Figure 1 A schematic flowchart of the channel-aware occupational exposure assessment method proposed in an embodiment of this disclosure is shown. Figure 2 This diagram illustrates the process of constructing a capability requirement vector for a target occupation according to an embodiment of this disclosure. Figure 3 A structural block diagram of a channel-sensing occupational exposure assessment device according to an embodiment of the present disclosure is shown. Figure 4 A simplified block diagram is shown for an electronic device 400 suitable for implementing embodiments of the present disclosure. Detailed Implementation

[0018] The present disclosure is described below based on embodiments, but it is not limited to these embodiments. In the detailed description of the present disclosure below, certain specific details are described in detail. However, those skilled in the art will fully understand the present disclosure without these details. To avoid obscuring the substance of the present disclosure, well-known methods, processes, flows, elements, and circuits are not described in detail.

[0019] Furthermore, those skilled in the art should understand that the accompanying drawings provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0020] Unless the context explicitly requires it, words such as "including" or "contains" throughout the application should be interpreted as including rather than exclusive or exhaustive; that is, meaning "including but not limited to".

[0021] In the description of this disclosure, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this disclosure, unless otherwise stated, "a plurality of" means two or more.

[0022] However, the inventors of this disclosure have discovered through research that the aforementioned prior art has the following technical defects: First, single-axis synthesis masks the heterogeneous mechanisms of action between the two technologies. Artificial intelligence (software form, near-zero marginal cost, diffusion cycle measured in months) and robotics (hardware form, high unit cost, diffusion cycle measured in years) have completely different dimensions, timelines, and capital structures in their impact on occupations. A single synthetic score cannot distinguish between "cognitive task exposure" and "physical task exposure," leading to systemic misjudgments in downstream decisions (such as retraining investment direction and workforce planning cycles). The inventors of this disclosure, by simultaneously comparing 516 standard occupational classifications (NOCs) samples with both indices, found that: the Frey index has a correlation coefficient as high as +0.71 with the robotics axis score of this disclosure, while its correlation with the AI ​​axis score is only +0.02—meaning that its so-called "automation" index is actually a "robotics / routine task" index; the Felten index has a correlation coefficient of +0.53 with the AI ​​axis score, while its correlation coefficient with the robotics axis score is only +0.02. 0.68 (significantly negative correlation). These two indices have long been used by academia and industry as interchangeable measures of "automatic exposure," and empirical studies have shown that they measure two nearly orthogonal or even inverse dimensions. Current technology has neither discovered this inversion phenomenon nor provided any method to systematically verify whether "multi-axis indices truly measure multiple separable constructs."

[0023] Second, the additive separability assumption systematically overestimates the importance of tightly coupled occupations. Tightly coupled occupations refer to those where cognitive judgment and real-time sensorimotor feedback are mutually dependent and inseparable in completing tasks, such as surgery, electrical work, and field nursing. The additive separability assumption breaks down the capabilities of these occupations into independent dimensions for comparison with cutting-edge technologies. This leads to the misjudgment that "each dimension can be accessed by technology" is equivalent to "the whole can be replaced," ignoring the fact that the real-time coupling between dimensions is precisely what machines struggle to replicate. Even when existing technologies acknowledge the coexistence of cognition and physical labor, they often approximate coupling using a product of "cognitive magnitude × sensorimotor magnitude." This approximation misjudges situations where "both cognitive and physical demands are high but occur at different times" as high coupling, failing to distinguish between the fundamentally different situations of "real-time bidirectional integration" and "co-occurrence of demands."

[0024] Third, existing methods mostly use a single scalar to uniformly represent automation risks, failing to provide differentiated assessments for two different types of automated entities: software artificial intelligence and embodied robots. Their scoring also largely relies on a single model, making cross-model analysis neither reproducible nor auditable. Furthermore, existing methods often construct occupational competency demand vectors directly from raw attributes or single sources, without cross-machining or interpretable projection of construct dimensions, nor adaptively weighting the reliability of multi-source fusion at the dimensional level. This results in insufficient interpretability and auditability of the demand vectors themselves.

[0025] Based on the findings of the above-mentioned problems, this disclosure provides a channel-aware occupational exposure assessment method. The method includes: acquiring multiple task-level natural language description texts of the target occupation; for each task-level natural language description text, integrating multiple large language models to obtain model output results, the model output results including: coupling strength, channel weight distribution, and a pair of bound dominant cognitive and dominant sensorimotor dimensions; wherein, the coupling strength is used to characterize the degree to which cognitive judgments of the corresponding task are instantly constituted through real-time sensorimotor feedback, the channel weight distribution includes the channel weights corresponding to the corresponding task on multiple predefined perceptual-motor channels; the bound dominant cognitive and dominant sensorimotor dimensions are the predetermined cognitive and sensorimotor dimensions corresponding to when cognitive judgments of the corresponding task can be instantly constituted through real-time sensorimotor feedback. The decoupling margin represents the proportion of tasks with zero coupling strength for the target occupation. Two independent capability frontier vectors are constructed to represent the current capability levels of the software AI and the embodied robot in different cognitive and sensorimotor dimensions, respectively. A basic exposure score for the target occupation is calculated based on the capability requirement vector of the target occupation and the two capability frontier vectors. The capability requirement vector represents the requirement level of the target occupation in different predetermined cognitive and sensorimotor dimensions. The basic exposure score represents the coverage of the capability levels of the software AI and the embodied robot in meeting the requirement level of the target occupation. The basic exposure score is adjusted by channel-by-channel discounting based on the channel weight distribution and the coupling strength to obtain the adjusted exposure score. This scheme calculates the coupling between cognitive judgment and real-time sensorimotor feedback from occupational task text, and uses the channel weight distribution to represent the perceptual-motor channel where the coupling occurs. Based on this, the automation exposure score is corrected using channel perception, enabling differentiated exposure assessment results for two different types of automated subjects: software AI and embodied robots. This disclosure can be applied to scenarios such as workforce planning, education investment decisions, personal career development guidance, and retraining resource allocation.

[0026] Based on the above description, this disclosure proposes a channel-aware occupational exposure assessment method, see [link to relevant documentation]. Figure 1 As shown, the method includes the following steps: S1. Obtain multiple task-level natural language description texts for the target occupation. In some embodiments, the task-level natural language description texts may be natural language descriptions of each job activity or responsibility of the target occupation from an authoritative occupation database. For example, the natural language descriptions of the target occupation may come from... The task statements and main duties descriptions from the Canadian National Occupational Classification (NOC) are used. Each target occupation can correspond to multiple task-level natural language description texts, for example, 5 to 30. In some embodiments, approximately 18,796 task-level natural language description texts can be obtained for 923 target occupations.

[0027] S2. For each task-level natural language description text, the model output is obtained by integrating multiple large language models. These multiple large language models can include N heterogeneous large language models, where N is an integer not less than two. Each large language model comes from an independent training institution or training source. The model output includes coupling strength, channel weight distribution, and a pair of bound dominant cognitive and dominant sensorimotor dimensions. It should be noted that the model output is obtained by integrating the outputs of each large language model according to a predefined scoring specification, and then using a weighted fusion method to obtain the final model output. Specifically, for the pair of bound dominant cognitive and dominant sensorimotor dimensions, each language model can directly identify and determine the pair according to the corresponding content in the predefined scoring specification; the final determination result is the integrated result of the various language models. For example, a large language model can directly determine which of the multiple cognitive dimensions is coupled in real time with which of the multiple sensorimotor dimensions within the closed loop of the real-time coupling of the task (i.e., the coupling of cognitive judgments required for the target profession under the task) in the target profession. The cognitive judgments involved in the task are then instantly linked to this effect through real-time sensorimotor feedback. The meanings of the model's output are as follows: (a) Coupling strength, used to characterize the degree to which cognitive judgments for a given task are formed instantly through real-time sensorimotor feedback, can take integer values ​​from 0 to 3. A coupling strength of 0 indicates no coupling, meaning the cognitive judgments for that task are not formed instantly through real-time sensorimotor feedback; that is, no cognitive dimension of the target occupation under that task is formed instantly through any real-time sensorimotor feedback. If the coupling strength is the highest, such as 3, then the cognitive judgments for that task are formed instantly through real-time sensorimotor feedback; that is, all cognitive dimensions of the target occupation under that task are formed instantly through the corresponding real-time sensorimotor feedback. In other words, coupling strength is determined using a "real-time bidirectional integration" approach, rather than a demand co-occurrence approach: only when cognitive judgments rely on feedback from any dimension such as the body or senses to be formed or corrected instantly is it considered high coupling strength (coupling strength = 3, such as surgical suturing or troubleshooting living circuits); if the cognitive judgments for the target occupation under that task and the demand for any sensorimotor dimension such as physical strength or senses both exist but occur at different times (such as designing before delivering code, moving building materials during the day, and studying construction drawings at night), it is considered low coupling strength or no coupling. It should be noted that when the coupling strength is zero, the channel weight distribution is empty.

[0028] (b) Channel weight distribution, including the channel weights corresponding to the task on multiple predefined perceptual-motor channels. In some embodiments, the channel weight distribution includes the channel weights of one dominant channel and at most two secondary channels selected from a predefined set of perceptual-motor channels. The channel weights of each channel in the channel weight distribution are located in the range [0,1], and the sum of the channel weights is 1. In some embodiments, the predefined perceptual-motor channels include six channels: visual-fine manipulation, tactile manipulation, proprioception and full-body posture, auditory judgment, real-time interpersonal interaction, and multi-sensory environment monitoring. The distinction between different channels is based on the dominant modality (visual, auditory, tactile, interpersonal, proprioception, etc.) that provides real-time feedback. For example, real-time hand-eye correction is visual-fine manipulation; force or texture feedback is tactile manipulation; body posture is proprioception and full-body posture; real-time sound judgment is auditory judgment; real-time response to people is real-time interpersonal interaction; and simultaneous monitoring of multiple environmental signals is multi-sensory environment monitoring. Channel weights are determined by integrating multiple large language models according to the corresponding content in the predefined scoring specifications: which sensorimotor dimension constitutes the cognitive judgment of the task in real time; the dominant channel corresponding to the dominant sensorimotor dimension is given the highest weight, and the secondary channel corresponding to the secondary sensorimotor dimension is given a secondary weight.

[0029] (c) A pair of bound dominant cognitive dimensions and dominant sensorimotor dimensions. The bound dominant cognitive dimension and dominant sensorimotor dimension are the predetermined cognitive dimension and predetermined sensorimotor dimension corresponding to the time when the cognitive judgment of the corresponding task can be instantly constituted through real-time sensorimotor feedback. Specifically, the bound sensorimotor dimension is the sensory or motor pathway that provides real-time feedback, and the bound cognitive dimension is the type of cognitive judgment that is instantly constituted or modified by that real-time feedback. In some embodiments, the dominant cognitive dimension may be selected from seven cognitive dimensions, including verbal, fluid reasoning, quantity, spatial, working memory, speed attention, and creativity; the dominant sensorimotor dimension may be selected from eleven sensorimotor dimensions, including fine manipulation, control coordination, reaction speed, near vision, far vision, color vision, hearing, strength, endurance, flexibility, and gross motor coordination. For example, the bound dominant cognitive dimension for surgical suturing is spatial, and the bound dominant sensorimotor dimension is fine manipulation; the bound dominant cognitive dimension for on-site circuit troubleshooting is fluid reasoning, and the bound dominant sensorimotor dimension is fine manipulation. In some embodiments, the multiple large language models can also integrate and output judgment reasons for subsequent verification.

[0030] It should be noted that the predefined scoring specifications in this disclosure include a set of cognitive dimensions, a set of sensorimotor dimensions, a set of channels, coupling strength anchor points, judgment criteria, and output patterns. These can be designed and fixed in advance according to actual needs, and the large language model cannot be changed.

[0031] S3. Obtain the coupling structure of the target occupation based on the coupling strength corresponding to the task and the channel weight distribution. The coupling structure includes total coupling strength, channel spectrum, and decoupling margin. Specifically: the total coupling strength is obtained by weighting the coupling strengths of multiple tasks. In some embodiments, the total coupling strength is a weighted average of the coupling strengths of each task, weighted by task importance: Σ_k w_k · strength_k, where w_k is the normalized weight of the task importance of the k-th task, and strength_k is the coupling strength of the k-th task. The channel spectrum is obtained by weighting the channel weight distributions of multiple tasks. In some embodiments, a multi-dimensional channel distribution vector of the target occupation can be obtained by weighting the channel weight distributions of each task by task importance and calculating the expectation: Σ_k w_k · dist_k, where dist_k is the channel weight distribution of the k-th task. The decoupling margin represents the proportion of tasks with zero coupling strength (i.e., no coupling) among all tasks for the target occupation. For example, it represents the weighted proportion of tasks with zero coupling strength, characterizing the weighted proportion of tasks in the target occupation that can be broken down into tasks to be planned and then executed. In some embodiments, the task importance is taken as... The normalized value of the task importance or relevance rating, with the sum of the task importance ratings being one.

[0032] S4. Construct two independent capability frontier vectors to represent the current capability levels achieved by software artificial intelligence and embodied robots in different cognitive and sensorimotor dimensions, respectively. The dimensions of the two capability frontier vectors are aligned with the capability dimensions of the capability requirement vector, which is equivalent to assigning a [0,1] frontier achievement level to each capability dimension. This frontier achievement level is obtained by normalizing the publicly available technical benchmark corresponding to that capability dimension. For example, the frontier achievement levels corresponding to software artificial intelligence include: fluid reasoning corresponding to GPQA (General Principles Q&A), working memory corresponding to long-chain reasoning benchmark, autonomy corresponding to SWE-bench (Software Engineering Benchmark) and METR (Center for Model Evaluation, Transparency, and Responsibility) task duration; the frontier achievement levels corresponding to embodied robots include: fine manipulation corresponding to FMB (Fine Manipulation Benchmark), gross coordination corresponding to Mobile-ALOHA (Mobile Aloha System), and whole-body corresponding to HumanoidBench (Humanoid Robot Benchmark). The two capability frontier vectors are constructed independently, allowing the same target occupation to be reached to different degrees by software artificial intelligence and by embodied robots.

[0033] In some implementations, the two capability frontier vectors satisfy axial purity constraints, meaning the product of the corresponding values ​​of the same capability dimension in the two frontier vectors is lower than a preset axial purity threshold (e.g., 0.05), ensuring that the two axes are structurally separable and avoiding two noisy versions of a single construct. In some embodiments, the sensorimotor dimension components of the capability frontier vector corresponding to the embodied robot are labeled according to the degree of structuring of the sensorimotor channel and the work context, forming a channel × context component set, wherein the components of open unstructured contexts are labeled independently of the components of structured contexts—thereby avoiding misjudging structured tasks (such as forklift handling) and open tasks (such as surgery) as equivalent substitutes. The sensorimotor dimension components of the capability frontier vector corresponding to the software artificial intelligence characterize its perception and control level achievable through the digital interface, taking the lowest value when there is no embodied execution capability. The capability frontier vector can be updated as the evaluation time changes to reflect the dynamic changes of the technological frontier.

[0034] S5. Calculate the basic exposure score for the target occupation based on the capability requirement vector and the two capability frontier vectors. The capability requirement vector characterizes the requirement level of the target occupation in different predetermined cognitive and sensorimotor dimensions; the basic exposure score characterizes the coverage of the capability levels of the software artificial intelligence and the embodied robot in meeting the requirement level of the target occupation. The basic exposure score is calculated by a coverage function: comparing the requirement level in the capability requirement vector with the frontier achievement level in the capability frontier vector dimension by dimension. The frontier achievement level indicates that the capability dimension that can reach or exceed the requirement level is counted as covered. Then, the coverage ratio is obtained by weighting the results of whether each capability dimension is covered, and normalized to 0 to 100, thus obtaining the basic exposure score. The basic exposure scores for the software artificial intelligence and the embodied robot are calculated separately.

[0035] S6. Based on the channel weight distribution and the coupling strength, the basic exposure score is adjusted by discounting each channel to obtain the adjusted exposure score.

[0036] In some embodiments, the channel discount adjustment of the base exposure score includes: applying a discount weighted adjustment to the base exposure score using a comprehensive weight obtained from the channel weights and total coupling strength in the channel spectrum; in some embodiments, when calculating the comprehensive weight from the channel weights and total coupling strength in the channel spectrum, the channel discount weight is also used for weighting. For example, the adjusted exposure score is equal to the base exposure score multiplied by a factor minus the sum of the products of the discount weights of each channel in the channel spectrum and the coupling strength component of that channel, i.e.: Adjusted = Base × (1 Σ_c α_c ·ECI_c), where α_c is the discount weight of channel c, and ECI_c is the coupling strength component of channel c. The coupling strength component ECI_c of channel c is equal to the weight component value of channel c in the channel spectrum multiplied by the normalized total coupling strength, representing the real-time coupling content of the target occupation on channel c. The discount weight α_c can be calibrated by employment or retention outcome data. The calibration method can be, for example, regression fitting of α_c to the actual retention or non-displacement rate of historically high c-channel coupling occupations. In some embodiments, different channels have different discount weights for the software AI axis and the embodied robot axis: the real-time interpersonal interaction channel applies a stronger discount to both the software AI axis and the embodied robot axis, that is, the channel discount weight corresponding to this channel is larger, and the vision-fine manipulation channel applies a greater channel discount weight to the embodied robot axis than to the software AI axis. The higher the coupling strength and the higher the channel discount weight, the lower the corresponding adjusted exposure score is adjusted—this is the exposure correction for tightly coupled occupations.

[0037] S7. Calculate the consistency among the outputs of the multiple large language models, and re-evaluate tasks with consistency below a threshold. The consistency includes at least one of the following three categories: consistency of channel weight distribution, consistency of coupling strength, and consistency between the bound dominant cognitive dimension and the dominant sensorimotor dimension. In some embodiments, coupling strength consistency is measured by a weighted kappa, Krippendorff α coefficient, or intragroup correlation coefficient (ICC); channel weight distribution consistency is measured by the cosine similarity or distribution distance between the channel weight distributions output by each model, rather than an exact match of the dominant channel category; bound dimension consistency is measured by the proportion of consistency between the bound dominant cognitive dimension and the dominant sensorimotor dimension pairs output by each model. Channel consistency is measured by distribution similarity rather than exact matching because forced single-choice might misclassify both models as hand-vision, only with different subdivisions, as inconsistent; this cosine similarity distribution comparison method improves consistency from approximately 79% for exact matching to approximately 90%. In some embodiments, when any consistency is below a preset threshold, the task is marked as a low consistency task and sent to a re-evaluation queue for re-evaluation by a human or a higher-specification model.

[0038] The construction details of the capability requirement vector are described below.

[0039] The competency requirement vector for the target occupation is constructed using the following steps, see below. Figure 2 As shown.

[0040] Step 201: Obtain quantitative scores for the target occupation across multiple predetermined ability dimensions from a pre-stored database, such as a professional ability survey database. The professional ability survey database is an authoritative professional ability survey database containing 52 ability dimensions (e.g., [database name missing]). The system comprises 21 dimensions of cognitive ability, 10 dimensions of psychomotor ability, 9 dimensions of physical ability, and 12 dimensions of sensory ability. The score for each ability dimension includes at least one of a level value and a importance value; the quantitative scoring uses either a level value or the product of a level value and an importance value.

[0041] Step 202 (Cognitive Sub-component Step): The scores of the cognitive ability dimensions in the quantitative scoring are projected onto the cognitive ability mapping sub-matrix of the predetermined mapping matrix to obtain the cognitive sub-components of the target occupation in D predetermined cognitive dimensions (i.e., the first set dimension). Taking the occupational ability survey database containing 52 ability dimensions as an example, 21 cognitive ability dimensions are projected onto the D predetermined cognitive dimensions.

[0042] The D predetermined cognitive dimensions are set a priori based on the cognitive ability factor theory. In one embodiment, D=7, which are verbal, fluid reasoning, quantity, spatial, working memory, speed and attention, and creativity.

[0043] The mapping matrix is ​​a sparse matrix, where each capability dimension has a non-zero weight only with at most two predetermined cognitive dimensions. These at most two predetermined cognitive dimensions include a primary attribution dimension (primary mapping edge) with a larger weight and a secondary attribution dimension (secondary mapping edge) with a smaller weight; and the sum of the non-zero weights of the same capability dimension across all predetermined cognitive dimensions is one (i.e., the sum of outgoing edge weights is equal to one). This is because mapping the same capability dimension simultaneously to two predetermined cognitive dimensions at a weight of 1.0 will simultaneously amplify the target occupation containing that capability across the corresponding predetermined cognitive dimensions, leading to distortion in cross-occupation comparisons.

[0044] The following table shows three typical types of mapping edges (taken from a 52×D table example):

[0045] The projection calculation involves two steps: First, collecting votes (including scores multiplied by edge weights) from each predetermined cognitive dimension; second, dividing by the sum of the incoming edge weights of each predetermined cognitive dimension to restore the weighted average on a scale of 0 to 7. For example, consider a simplified profession (such as accountant), assuming its... The scoring only involves four abilities: verbal comprehension 4.0 (outbound edge C1 × 1.0), inductive reasoning 5.0 (outbound edge C2 × 1.0), mathematical reasoning 6.0 (outbound edge C3 × 0.7, C2 × 0.3), and calculation ability 6.5 (outbound edge C3 × 1.0). Therefore: C1 = 4.0 × 1.0 = 4.00, with an incoming edge weight sum of 1.0, resulting in 4.00; C2 = 5.0 × 1.0 + 6.0 × 0.3 = 6.80, with an incoming edge weight sum of 1.3, resulting in 5.23; C3 = 6.0 × 0.7 + 6.5 × 1.0 = 10.70, with an incoming edge weight sum of 1.7, resulting in 6.29. The cognitive subcomponent (cognitive part) = [C1 = 4.00, C2 = 5.23, C3 = 6.29, ...], has the strongest quantity, aligning with accounting intuition. Interpretability lies in the fact that C3=6.29 can be traced back to being driven by mathematical reasoning (0.7) and computational ability (1.0), and the score of each ability dimension can be traced back to specific factors. Capabilities and their weights. Cognitive subcomponents.

[0046] Step 203 (Model Output Sub-component Step): Input the occupational description text of the same target occupation into at least three independent large language models to obtain the demand scores of each model for the target occupation on the D predetermined cognitive dimensions. Then, perform a trust-weighted ensemble on the scores of each model to obtain the model output sub-component. The trust weight of each model is pre-calibrated based on its historical consistency with the anchoring benchmark. Before integration, the scores of each model are centered dimension-wise by model to eliminate systematic bias between models. The model output sub-component can supplement the fine granularity lost in dimensionality reduction, that is, reducing the multidimensional cognitive ability assessment information to only the first set dimension of the cognitive sub-component.

[0047] Step 204 (Sensory-Motor Requirement Component Step): Using the sensorimotor ability mapping submatrix of the mapping matrix, the scores of the sensorimotor ability dimensions in the quantitative scores of the multiple predetermined ability dimensions are projected to obtain the sensorimotor requirement components of the target occupation in multiple predetermined sensorimotor dimensions. Taking the occupational ability survey database containing 52 ability dimensions as an example, the 31 ability dimensions of the other categories (excluding the 21 cognitive abilities) are projected onto the predetermined sensorimotor dimensions to obtain the sensorimotor requirement components. The predetermined sensorimotor dimensions include multiple of the following: fine motor skills, control precision, reaction and multi-limb coordination, strength, endurance and flexibility, gross coordination and balance, near vision, far vision, and hearing. The sensorimotor requirement components are obtained solely through mapping (single-path), parallel to the fusion in Step 205, and not involved in cognitive fusion.

[0048] Step 205 (Fusion Step): The cognitive sub-component and the model output sub-component are fused according to their dimensions using a reliability-weighted method to obtain the cognitive demand component. The fusion weight α_d of the d-th dimension is set based on the cross-path consistency between the cognitive sub-component and the model output sub-component on the sample occupation set: dimensions with high consistency increase the weight of the cognitive sub-component, and dimensions with low consistency increase the weight of the model output sub-component. Specifically, the fusion weight α_d of the d-th dimension is: α_d = clip(a + b·(r_d) r ), α_min, α_max), where r_d is the correlation coefficient between the cognitive component and the model output component on the sample occupation set in this dimension, r The α_d represents the mean correlation coefficients across all dimensions, and α_clip is the clipping function; in one embodiment, a = 0.5, b = 0.6, α_min = 0.35, and α_max = 0.75. This dimension-level α_d (not a globally singular α) achieves adaptive reliability: which path is more reliable in which dimension is determined by the data. During fusion, for target occupations where the standardized difference between the cognitive sub-component and the model output sub-component exceeds a predetermined threshold in any dimension, a divergence marker is added and output for verification; in one embodiment, the predetermined threshold can be a standardized difference of 2.

[0049] Step 206 (Spoken Assembly): The cognitive requirement component and the sensorimotor requirement component are spliced ​​together to form the capability requirement vector for the target occupation, i.e., capability requirement vector = [cognitive requirement component || sensorimotor requirement component]. This method ensures that the 31 sensorimotor dimensions are not discarded or contaminated by the cognitive dimension; the cognitive and sensorimotor dimensions are independently normalized, and the coordinate system semantics are clear.

[0050] In one embodiment, a pilot study was conducted with approximately 499 occupations (bridged to a survey database using an occupational classification lookup table). The correlation coefficients between the cognitive subcomponent and the model output subcomponent were as follows: verbal 0.885, fluid reasoning 0.864, quantity 0.860, spatial 0.692, working memory 0.796, speed and attention 0.534, and creativity 0.897, with a mean of 0.790. Using the aforementioned fusion weighting formula, α = {verbal 0.557, fluid reasoning 0.545, quantity 0.542, spatial 0.441, working memory 0.504, speed and attention 0.350, creativity 0.564} – the speed and attention dimension showed the lowest consistency, automatically biasing towards the second path. In a divergence identification example, when the threshold was set to a standardized difference of 2, 33 (6.6%) of the 499 target occupations were labeled, concentrated in the spatial or speed dimensions of performing and first-aid occupations, consistent with the prior diagnosis that "time pressure and physical expression needs are underestimated due to dimensionality reduction."

[0051] In another embodiment, the scores of at least three large language models in the model output sub-component step can be replaced by the output of a single model distilled using multi-model consensus annotations as the training target. The distilled model maintains the same dimensional set and scoring caliber as the replaced multi-model ensemble. A single distilled model is trained with approximately 62,000 multi-model consensus annotations. On the holdout set: the average correlation of cognitive D-dimensional data is 0.890, the cosine distribution of sensorimotor dimensions is 0.937, the coupling strength consistency is 99.2%, and the structured output parsing failure rate is 0. This single model can replace multi-model ensembles for real-time decomposition, supporting online real-time construction of demand vectors for each profession or task.

[0052] The beneficial effects of this disclosure are as follows: First, by extracting coupling as "strength-channel weight distribution-bound dimension" and aggregating it into a channel spectrum, embodied coupling is upgraded from a single scalar to a structured description. This allows for the application of differentiated exposure discounts to software AI and embodied robots separately, overcoming the systematic overestimation of tightly coupled occupations by the additive separability assumption. The channel spectrum (vector) rather than scalar representation of coupling strength allows the coupling strength to be distributed according to predefined channels, characterizing the coupling structure. The channel spectrum applies different channel weight discounts to software AI and embodied robots: the interpersonal channel strongly discounts both software AI and embodied robots, while the visual-fine channel primarily discounts the embodied robot side, thus providing structured and differentiated exposure corrections for the two types of automated agents.

[0053] Second, the scoring standard uses real-time bidirectional integration rather than demand co-occurrence as the coupling criterion, avoiding misjudging occupations where both cognitive and physical demands are high but occur at different times as having high coupling. This is a fundamental difference from the existing approximation method of the product of cognitive magnitude and sensory magnitude, making the extraction of coupling strength more consistent with embodied cognition theory.

[0054] Third, replacing a single dominant channel with channel weight distribution and using distribution similarity to measure consistency eliminates the spurious discrepancies created by forced single selection across models, significantly improving the reproducibility and auditability across heterogeneous models. Experimental results show that for approximately 18,800 tasks, the scores of two heterogeneous models from different vendors have the following characteristics: a coupling strength correlation coefficient of approximately 0.87, with approximately 98% of differences not exceeding 1; a cosine similarity of approximately 0.90 between channel weight distributions, and a dominant consistency rate of approximately 95% at the four cluster levels with reliable resolution—a significant improvement compared to approximately 79% when channels were forced into a single category and measured by exact matching in earlier methods.

[0055] Fourth, the cognitive requirement component of the capability requirement vector is constructed using a dual-path complementary and auditable design: the first path (mapping matrix projection) is objectively interpretable, with each dimension allowing for the lookup of the driving capability dimension and its weights; the second path (multi-model integration) supplements the fine-grained details lost during dimensionality reduction; the two are fused by dimension-level reliability weights, rather than a single global weight. The sensorimotor component is not discarded or mixed in dimensions: parallel branches provide faithful projection, and the cognitive and sensorimotor sets are respectively normalized, resulting in a clear semantic coordinate system.

[0056] Fifth, the frontier is well-defined: it is set in the same dimension, anchored at the same scale, and can be updated at specified time points; the frontier of the embodied robot is calibrated by channel × context, avoiding systematic exposure misjudgments caused by the confusion between structured and open contexts. By using a constant capability requirement vector and aggregation formula parameters, and only updating the capability frontier vector at several future time points, the validity verification of the baseline time point remains effective across the entire projected time domain, and supports cross-version differential auditing.

[0057] Sixth, by superimposing a social discount based on stake dimension ratings on top of the channel discount, we can capture the irreplaceability that cannot be represented by either the capability axis or the coupling axis, stemming from whether or not it is suitable for machines to perform. For example, professions such as judges and court interpreters have high exposure to both capabilities and coupling (cognition can be accessed by technology and is not tightly coupled), but their stake content is high (responsibility, legality, and presence are significant). After the stake layer discount, the exposure decreases significantly, revealing the irreplaceability that cannot be captured by either the capability axis or the coupling axis, but can only be captured by the stake layer.

[0058] The following describes some technical details of this disclosure through exemplary embodiments.

[0059] Example 1: Task-level Coupled Annotation and Aggregation This embodiment exemplifies the detailed implementation of steps S1 to S3. First, multiple task-level natural language description texts of the target occupation are obtained from the task database. These task-level natural language description texts are natural language descriptions of each work activity or responsibility of the occupation from an authoritative occupation database. The data source is... The task statements are compared to the main duties of the Canadian National Commission (NOC). For example, several task-level natural language descriptions for a surgeon include: "Operate on patients to remove, repair, or improve functioning of diseased or injured body parts" and "Examine instruments, equipment, and operating room to ensure sterility," etc. A typical occupation has 5 to 30 such statements; in this disclosure example, There are approximately 923 occupations and a total of about 18,796 task statements.

[0060] For each task-level natural language description text, a model panel consisting of N heterogeneous large language models (N is an integer greater than or equal to 2) is scheduled, and a quadruple {coupling strength, channel weight distribution, bound cognitive dimension, bound sensory-motor dimension} is output according to a predefined scoring specification. The predefined scoring specification is a predefined structured scoring guide (prompt / rubric). It is important to emphasize that the bound cognitive dimension and bound sensory-motor dimension are directly determined by the model, rather than by assigning the highest score to each dimension—the model directly determines "which cognitive dimension and which sensory-motor dimension are bound together in real time in the real-time coupling loop of this task".

[0061] The coupling strength takes integer values ​​from 0 to 3, with the following anchor points: 0 = cognition and physical or sensory processes occur separately at different times (e.g., "design before code delivery"); 1 = weak coupling; 2 = medium coupling; 3 = cognitive judgment must be bidirectionally constituted by real-time sensory-motor feedback and cannot be separated (e.g., surgical suturing, troubleshooting in living circuits). The core criterion is real-time bidirectional integration rather than demand co-occurrence; that is, if both cognitive and physical demands are high but occur at different times, it is judged as low coupling. The channel weight distribution includes the weights of one dominant channel and at most two secondary channels from six predefined perceptual-motor channels. The weights of this dominant channel and at most two secondary channels are between [0,1] and are set to 1. When the coupling strength is zero, the channel weight distribution is empty.

[0062] Taking surgical suturing as an example: Coupling strength = 3 (real-time bidirectional hand-eye integration), channel weight distribution = {visual-fine manipulation: 0.6, tactile manipulation: 0.4}, bound dominant cognitive dimension = spatial (C4), bound dominant sensorimotor dimension = fine manipulation. Taking on-site circuit troubleshooting as an example: Coupling strength = 3, channel weight distribution = {visual-fine manipulation: 0.5, tactile manipulation: 0.3, auditory judgment: 0.2}, bound dominant cognitive dimension = fluid reasoning (C2), bound dominant sensorimotor dimension = fine manipulation. Taking programmers writing code as an example: Coupling strength = 0 (think before you act), channel weight distribution is empty, no bound dimension.

[0063] Then, using task importance as the weight, the coupling strength of each task and the channel weight distribution are aggregated into a professional-level coupling structure. Task importance is taken as... The task importance or relevance ratings are normalized, with the sum of the weights of each task being one. Total coupling strength = Σ_k w_k · strength_k; channel spectrum = Σ_k w_k · dist_k (6-dimensional distribution vector); decoupling margin = Σ_{k:strength_k=0}w_k. Taking a surgeon as an example, out of approximately 20 tasks, about 12 are highly coupled (strength 3), about 5 are moderately coupled (strength 1 to 2), and about 3 are uncoupled (strength 0), with a total coupling strength of approximately 2.50. The channel spectrum is mainly visual-fine manipulation and tactile manipulation, with a decoupling margin of approximately 0.15.

[0064] Example 2: Channel-Aware Exposure Discount This embodiment exemplifies the detailed implementation of steps S4 to S6 above. First, two independent capability frontier vectors are constructed, representing the software artificial intelligence capability domain and the embodied robot capability domain, respectively. For each capability frontier vector, a [0,1] frontier achievement level is assigned to the capability dimension, which is obtained by normalization from the publicly available technical benchmarks corresponding to that capability dimension. For example, the frontier achievement levels on the software artificial intelligence side are: fluid reasoning corresponding to GPQA, working memory corresponding to long-chain reasoning benchmarks, and autonomy corresponding to SWE-bench and METR task duration; for example, the frontier achievement levels on the embodied robot side are: fine manipulation corresponding to FMB, gross coordination corresponding to Mobile-ALOHA, and whole-body corresponding to HumanoidBench.

[0065] The two capability frontier vectors satisfy the axis purity constraint: the product of the weights of the same capability dimension in the two capability frontier vectors is lower than a preset axis purity threshold (e.g., 0.05), ensuring that the two axes are structurally separable. The first frontier vector (software AI side) mainly weights cognitive and learning dimensions, such as verbal (0.18), mathematical logic (0.16), pattern recognition under uncertainty (0.14), and continuous learning (0.10), with weights approaching zero for psychomotor-physical and perceptual dimensions; the second frontier vector (embodied robot side) mainly weights psychomotor-physical and perceptual dimensions, such as fine motor skills (0.15), large muscle group movements (0.13), tactile-proprioception (0.12), visual perception (0.11), and strength and endurance (0.10), with weights approaching zero for cognitive and learning dimensions.

[0066] Based on the capability requirement vector of the target occupation and two capability frontier vectors, a base exposure score is calculated using a coverage function: The capability requirement level is compared dimension-by-dimensionally with the frontier achievement level; dimensions where the frontier meets or exceeds the requirement are considered covered. The coverage ratio is obtained by weighting the dimensions and normalizing to 0-100. The calculation is performed once each on the software AI side and the embodied robot side, yielding the base exposure score for the AI ​​axis and the base exposure score for the embodied robot axis. Subsequently, the base exposure score is discounted channel-by-channel using the channel spectrum. The formula for the adjusted exposure score is: Adjusted = Base × (1 Σ_c α_c ·ECI_c), where α_c is the discount weight of channel c (defined by employment or retention outcome data), and ECI_c is the coupling strength component of channel c (equal to the component value of channel c in the channel spectrum multiplied by the normalized total coupling strength).

[0067] Different channels apply different discount weights to the AI ​​and embodied robot axes: the real-time human-to-human interaction channel applies a strong discount to both the AI ​​and embodied robot axes (α_AI approximately 0.35, α_Robot approximately 0.30); the vision-fine manipulation channel primarily discounts the embodied robot axis (α_AI approximately 0.10, α_Robot approximately 0.40); and the haptic manipulation channel primarily discounts the embodied robot axis (α_AI approximately 0.05, α_Robot approximately 0.35). Therefore, the same target occupation receives differentiated adjusted exposure scores on the AI ​​and embodied robot axes. It should be noted that the method for calibrating based on employment or retention outcome data could be, for example, regression fitting of α_c to the actual retention or non-displacement rate of historically high c-channel coupling occupations. The stronger the coupling and the higher the channel discount weight, the lower the exposure is suppressed, which can achieve exposure correction for tightly coupled occupations.

[0068] Example 3: Multi-source fusion and temporal update of capability requirement vectors This embodiment exemplifies the construction details of the capability requirement vector, which is constructed by dual-path fusion: the first path is the projection of the mapping matrix, and the second path is the integration of multiple models.

[0069] First path (cognitive subcomponent): From Obtain the target occupation's scores across 52 ability dimensions (21 columns for cognitive, 10 columns for psychomotor, 9 columns for physical, and 12 columns for sensory). Project the cognitive ability dimension scores onto a 52×D weighted cross-mapping matrix to obtain D=7 cognitive sub-components across predetermined cognitive dimensions. This cognitive ability mapping matrix is ​​sparse; each ability dimension has non-zero weights on at most two predetermined cognitive dimensions (primary and secondary mapping edges), and the sum of outgoing edge weights for the same ability dimension is uniform. The projection calculation is as follows: calculate the score for each dimension × edge weight, then divide by the sum of incoming edge weights to obtain the weighted average. The score for each dimension can be used to look up the driving ability dimension and its weight.

[0070] The second approach (model output sub-component): The task-level natural language description text of the target profession is input into at least three independent large language models to obtain the demand scores of each model on D=7 cognitive dimensions. After centering by model dimension by dimension, the models are weighted and integrated to obtain the model output sub-component. The trust weights of each model are pre-calibrated based on their historical consistency with the anchoring benchmark. The model output sub-component supplements the fine-grained details lost in dimensionality reduction.

[0071] Sensorimotor demand component: The sensorimotor demand component is obtained by projecting the scores of the sensorimotor ability dimension onto the sensorimotor ability mapping submatrix of the mapping matrix. This component is a one-way projection.

[0072] Fusion: The cognitive sub-component and the model output sub-component are fused according to the dimensional reliability weight α_d to obtain the cognitive demand component. α_d = clip(a + b·(r_d)) r ), α_min, α_max), where r_d is the correlation coefficient between the cognitive sub-component and the model output sub-component on the sample occupation set; in one embodiment, a=0.5, b=0.6, α_min=0.35, α_max=0.75. Concatenation: The cognitive demand component and the sensorimotor demand component are concatenated to form the capability demand vector.

[0073] In one instance, for approximately 499 occupational pilots, the cognitive sub-component and the model output sub-component showed a dimension-wise correlation: verbal 0.885, fluid reasoning 0.864, quantity 0.860, spatial 0.692, working memory 0.796, speed and attention 0.534, and creativity 0.897, with a mean of 0.790. The fusion weight α = {verbal 0.557, fluid reasoning 0.545, quantity 0.542, spatial 0.441, working memory 0.504, speed and attention 0.350, creativity 0.564}, where the speed and attention dimension showed the lowest consistency, automatically biasing towards the second path. Divergence indicators: When the threshold was set to a standardized difference of 2, 33 (6.6%) of the 499 occupations were identified, concentrated in the spatial or speed dimensions of performing arts and first aid occupations.

[0074] Time-series update: The capability frontier vector can be updated at different evaluation points. Construct capability frontier vectors for the software AI and embodied robot at multiple future time points; repeat steps S5 to S6 to obtain the adjusted exposure scores for the target occupation at multiple future time points. By using a constant capability demand vector and aggregation formula parameters, and updating only the capability frontier vector, the primary validity verification at the baseline time point remains effective across the entire projected time domain, and supports cross-version differential auditing. Only the capability frontier vector is updated between versions; the formula form and weight structure are frozen, thus ensuring strict comparability across time domains and versions.

[0075] Example 4: Cross-model consistency gating (soft channel) This embodiment exemplifies the detailed implementation of step S7. First, the consistency among multiple large language model scores within the model panel is calculated. The consistency includes at least one of the following three categories: consistency of channel weight distribution, consistency of coupling strength, and consistency between the bound dominant cognitive dimension and the dominant sensorimotor dimension.

[0076] Coupling strength consistency: For each task, calculate the weighted kappa, Krippendorff α, or intragroup correlation coefficient (ICC) among the coupling strength scores output by the N large language models. Channel weight distribution consistency: Measured by the cosine similarity or distribution distance between the channel weight distributions output by the major language models. Distribution similarity is used instead of exact matching to measure channel consistency because forced single-choice might misclassify both models as hand-visual, only with different subdivisions, as inconsistent. Additionally, consistency of the dominant clusters can be compared at the four main clusters (hand-visual, interpersonal, auditory, proprioceptive) at reliable channel resolution, serving as a coarse-grained measure of channel consistency. Bound dimension consistency: Measured by the proportion of consistency between the bound dominant cognitive dimension and the dominant sensorimotor dimension pairs output by the major language models.

[0077] When the consistency falls below a preset threshold (e.g., weighted kappa < 0.6), the task is marked as a low-consistency task and sent to the review queue. Tasks in the review queue can be re-evaluated by humans or by a higher-specification model, and the review results are written back to the task database with the review source marked.

[0078] In one example, two heterogeneous large language models from different vendors scored approximately 18,800 tasks: the coupling strength correlation coefficient was approximately 0.87, and the proportion of differences not exceeding 1 was approximately 98%; the cosine similarity between channel weight distributions was approximately 0.90, and the dominant consistency rate at the four cluster levels with reliable resolution was approximately 95%, representing a significant improvement of 79% compared to earlier methods that forced channels to a single category and used exact matching as the metric. This example demonstrates that representing by distribution and measuring by distribution similarity can eliminate spurious discrepancies created by forced single-choice selection. Approximately 4% of tasks entered the re-evaluation queue due to consistency falling below the threshold, and consistency significantly improved after re-evaluation. This gating mechanism makes the evaluation results reproducible and auditable, avoiding systematic contamination of the evaluation results by a single model bias.

[0079] It should be noted that when a certain reasoning model is used as a third judge, it systematically biases towards a single visual channel in the structured channel determination. Therefore, in some embodiments, two non-reasoning heterogeneous models can be used as anchors and the reasoning model can only participate in the intensity scoring when measuring channel consistency. This further ensures the robustness of the channel consistency measurement.

[0080] Example 5: Social Discount Based on Stakeholder Level This embodiment exemplifies the social discounting mechanism at the stake level. Based on the adjusted exposure score obtained after channel discounting in step S6, the model panel further outputs stake content assessment information for each task across several stake dimensions. Stake content can be output in a scoring process parallel to step S2 (using the same model panel and a different set of stake specifications), without being mixed into the coupled quadruple. The stake dimensions include at least one of responsibility, presence, relationship, legitimacy, and origin, each taking an integer value from 0 to 3.

[0081] Using task importance as the weight, the stake scores of each task are aggregated into a professional-level stake score. The aggregation method is: Professional-level stake score = Σ_k w_k · stake_k, where the professional-level stake score is the weighted sum of the importance of the 5-dimensional stake scores of each task. Then, a stake-layer discount is applied to the adjusted exposure score to obtain the stake-corrected exposure score: Stake-corrected exposure score = Adjusted exposure score × (1 γ (occupational-level risk content), where γ is a risk discount factor, determined from employment or retention outcome data, using the same method as α_c. Specifically, a preset risk discount factor γ is obtained; the risk content assessment information is corrected using the risk discount factor γ to obtain the risk-corrected exposure score.

[0082] In one instance, professions such as judges and court interpreters exhibit high exposure to both competence and coupling (cognition is technologically accessible and not tightly coupled), but also high stake content (responsibility, legality, and significant presence). After stake discounting, the exposure decreases significantly—revealing an irreplaceable nature that cannot be captured by either the competence or coupling axes, but only by the stake layer. This design allows this disclosure to further capture, beyond the competence and coupling axes, the social level of irreplaceability stemming from whether or not machines are suitable for such tasks, providing a more comprehensive correction for the automated exposure assessment of high-stakes professions.

[0083] Example 6: Channels determined from data by clustering This embodiment exemplifies a channel clustering elimination scheme. After outputting channel weight distributions for a large number of tasks, the channel weight distributions corresponding to each task are used as samples to cluster the predefined multiple perception-action channels. Based on the clustering results, a set number of channel weights corresponding to the perception-action channels with the fewest clusters are eliminated from the channel weight distributions.

[0084] Specifically, using the channel weight distribution of all tasks as a sample, hierarchical clustering or k-means clustering is performed on six predefined channels. The order in which each channel is merged and the number of clusters are observed in the clustering tree. If a channel has extremely low weights in most tasks or is highly co-occurring with other channels, it will be the channel with the fewest clusters in the clustering—that is, this channel lacks independent support in the data and is a channel that the model cannot reliably distinguish. A set number of channels with the fewest clusters (e.g., 1 to 2) are removed, and the remaining channels are re-normalized as the effective channel set for subsequent use, such as for the aggregation of coupled structures and the channel-by-channel discount adjustment of the basic exposure score.

[0085] In one example, clustering of channel weight distribution for approximately 18,800 tasks revealed that five channels—visual-fine manipulation, tactile manipulation, proprioception and full-body posture, auditory judgment, and real-time interpersonal interaction—each had independent support. The multi-sensory environmental monitoring channel, however, showed high co-occurrence with other channels in most tasks and had the fewest clusters, thus it was removed. After removal, the remaining five channels were re-normalized, further improving channel consistency. This embodiment demonstrates that the channel set is not static but can be adaptively determined by the data, making this disclosure adaptable to different professional fields and task distributions.

[0086] Figure 3 This diagram illustrates a structural block diagram of a channel-aware occupational exposure assessment device according to an embodiment of the present disclosure. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 3 As shown, the channel-aware occupational exposure assessment device includes: The acquisition module 301 is configured to acquire multiple task-level natural language description texts of the target occupation; The model output module 302 is configured to obtain model output results by integrating multiple large language models for each task-level natural language description text. The model output results include: coupling strength, channel weight distribution, and a pair of bound dominant cognitive dimension and dominant sensorimotor dimension. The coupling strength is used to characterize the degree to which the cognitive judgment of the corresponding task is formed instantly through real-time sensorimotor feedback. The channel weight distribution includes the channel weights of the corresponding task on multiple predefined perceptual-motor channels. The bound dominant cognitive dimension and dominant sensorimotor dimension are the predetermined cognitive dimension and predetermined sensorimotor dimension, respectively, when the cognitive judgment of the corresponding task can be formed instantly through real-time sensorimotor feedback. The vector construction module 303 is configured to construct two independent capability front vectors, which respectively represent the current capability levels of the software artificial intelligence and the embodied robot in different cognitive and sensorimotor dimensions. The basic exposure module 304 is configured to calculate the basic exposure score of the target occupation based on the capability requirement vector of the target occupation and the two capability frontier vectors; the capability requirement vector is used to characterize the requirement level of the target occupation in different predetermined cognitive dimensions and predetermined sensorimotor dimensions; the basic exposure score is used to characterize the coverage of the capability level of the software artificial intelligence and the embodied robot in meeting the requirement level of the target occupation. The exposure adjustment module 305 is configured to perform a channel-by-channel discount adjustment on the basic exposure score based on the channel weight distribution and the coupling strength to obtain the adjusted exposure score. The technical features and their corresponding explanations and descriptions involved in the above-described device embodiments are the same as, corresponding to, or similar to the technical features and their corresponding explanations and descriptions involved in the above-described method embodiments. For the technical features and their corresponding explanations and descriptions involved in the above-described device embodiments, please refer to the technical features and their corresponding explanations and descriptions involved in the above-described method embodiments. This disclosure will not repeat them here.

[0087] Figure 4 A simplified block diagram is shown for an electronic device 400 suitable for implementing embodiments of the present disclosure. Figure 4 As shown, the electronic device 400 includes a processing unit 401, which can be implemented as a CPU, GPU, FPGA, NPU, or other processing unit. The processing unit 401 can execute various processes according to any of the methods described above in this disclosure, based on a program stored in the read-only memory (ROM) 402 or a program loaded from the storage portion 408 into the random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the computer system 400. The processing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0088] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0089] In particular, according to embodiments of this disclosure, any of the methods described above in the embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program containing program code for performing any of the methods in the embodiments of this disclosure. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411.

[0090] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, apparatus (devices), or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This disclosure is described with reference to flowchart illustrations of methods, apparatus (devices), and computer program products according to embodiments of this disclosure. It should be understood that each step in the flowchart can be implemented by computer program instructions.

[0092] These computer program instructions may be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction means, the implementation process of which is described in the instruction means. Figure 1 The function specified in one or more processes.

[0093] These computer program instructions may also be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, produce instructions for implementing processes. Figure 1 A device for a function specified in one or more processes.

[0094] Another embodiment of this disclosure relates to a non-volatile storage medium for storing a computer-readable program for use by a computer to execute some or all of the above-described method embodiments.

[0095] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by specifying relevant hardware through a program. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0096] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.

Claims

1. A channel-aware occupational exposure assessment method, characterized in that, The method includes the following steps: Obtain multiple task-level natural language description texts for the target profession; For each task-level natural language description text, multiple large language models are integrated to obtain model output results. The model output results include: coupling strength, channel weight distribution, and a pair of bound dominant cognitive dimension and dominant sensorimotor dimension. The coupling strength is used to characterize the degree to which the cognitive judgment of the corresponding task is formed instantly through real-time sensorimotor feedback. The channel weight distribution includes the channel weights of the corresponding task on multiple predefined perceptual-motor channels. The bound dominant cognitive dimension and dominant sensorimotor dimension are the predetermined cognitive dimension and predetermined sensorimotor dimension, respectively, when the cognitive judgment of the corresponding task can be formed instantly through real-time sensorimotor feedback. Two independent capability frontier vectors are constructed to represent the current capability levels of software artificial intelligence and embodied robots in different cognitive and sensorimotor dimensions, respectively. The basic exposure score of the target occupation is calculated based on the capability requirement vector of the target occupation and the two capability frontier vectors; the capability requirement vector is used to characterize the requirement level of the target occupation in different predetermined cognitive and sensorimotor dimensions; the basic exposure score is used to characterize the coverage of the capability level of the software artificial intelligence and the embodied robot in meeting the requirement level of the target occupation. The basic exposure score is adjusted by discounting each channel based on the channel weight distribution and the coupling strength to obtain the adjusted exposure score.

2. The method according to claim 1, characterized in that, The method further includes: The coupling structure of the target occupation is obtained based on the coupling strength corresponding to the task and the channel weight distribution. The coupling structure includes the total coupling strength obtained by weighting the coupling strength of multiple tasks, the channel spectrum obtained by weighting the channel weight distribution of multiple tasks respectively, and the decoupling margin. The decoupling margin represents the proportion of tasks with a coupling strength of 0 for the target occupation.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Using the channel weight distribution as a sample, cluster the predefined multiple perception-action channels, and based on the clustering results, remove a set number of perception-action channels and their corresponding channel weights that have the fewest clusters from the channel weight distribution.

4. The method according to claim 1 or 2, characterized in that, The method further includes: Constructing the future capability frontier vectors of software artificial intelligence and embodied robots at multiple points in the future; Based on the future capability frontier vector, the adjusted exposure score of the target profession is obtained at multiple future time points.

5. The method according to claim 1 or 2, characterized in that, The method further includes: Obtain multidimensional competency assessment information for the target occupation from a designated database; The multidimensional ability assessment information is projected onto multiple predetermined cognitive dimensions and multiple predetermined sensorimotor dimensions using a predetermined mapping matrix to obtain the ability requirement vector for the target occupation.

6. The method according to claim 5, characterized in that, The predetermined mapping matrix includes a cognitive ability mapping sub-matrix and a sensorimotor ability mapping sub-matrix; The multidimensional ability assessment information is projected onto multiple predetermined cognitive dimensions and multiple predetermined sensorimotor dimensions using a predetermined mapping matrix to obtain the ability requirement vector for the target occupation, including: The cognitive ability assessment information in the multidimensional ability assessment information is projected onto multiple predetermined cognitive dimensions using the cognitive ability mapping sub-matrix to obtain cognitive sub-components; The cognitive sub-component is weighted and fused with the model output sub-component to obtain the cognitive demand component; the model output sub-component is obtained by evaluating the task-level natural language description text of the target occupation on the predetermined cognitive dimension through multiple large language models. The sensorimotor ability assessment information in the multidimensional ability assessment information is projected onto the predetermined sensorimotor dimension using the sensorimotor ability mapping submatrix to obtain the sensorimotor demand component. The cognitive demand component and the sensorimotor demand component are concatenated to obtain the capability demand vector.

7. The method according to any one of claims 1, 2, and 6, characterized in that, The method further includes: Calculate the consistency among the outputs of the multiple large language models; the consistency includes the consistency of channel weight distribution, the consistency of coupling strength, and / or the consistency between the bound dominant cognitive dimension and the dominant sensorimotor dimension.

8. The method according to any one of claims 1, 2, and 6, characterized in that, The method further includes: For each task-level natural language description text, the multiple large language models are used to further obtain the interest content assessment information of the target occupation in multiple predefined interest dimensions; The adjusted exposure score is adjusted using the benefit / harm content assessment information to obtain the benefit / harm corrected exposure score.

9. The method according to claim 8, characterized in that, The adjusted exposure score is adjusted using the aforementioned benefit / risk content assessment information to obtain a benefit / risk-corrected exposure score, including: Obtain the preset benefit / discount factor; The adjusted exposure score is corrected using the interest discount factor to obtain the interest-corrected exposure score.

10. The method according to any one of claims 1, 2, 6, and 8, characterized in that, The dominant cognitive dimensions include at least one of the following: verbal, fluid reasoning, quantity, spatial, working memory, speed attention, and creativity; the dominant sensorimotor dimensions include at least one of the following: fine manipulation, control coordination, reaction speed, near vision, far vision, color vision, hearing, strength, endurance, flexibility, and gross motor coordination.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-10.

12. An electronic device comprising a memory and a processor, characterized in that, When the processor executes the computer program in the memory, it implements the steps of the method according to any one of claims 1-10.

Images