A hierarchical attention-based large-scale model inference system and method for complex decision-making.

Abstract

This invention relates to the fields of artificial intelligence and reasoning technology, and discloses a hierarchical attention-based large-scale model reasoning system and method for complex decision-making. The system includes a target intent module; a path planning module; an operation verification module; a potential energy processing module; and a reasoning decision-making module. The method includes: receiving a query, decomposing it into a target kernel and constructing an ideal potential energy profile; generating multiple parallel reasoning paths based on this profile; aggregating each path into a single polarized potential energy vector through evidence retrieval and quantification; calculating the potential energy alignment of each path and monitoring whether its risk component exceeds a safety threshold; if it exceeds the threshold, generating a hedging subtask to adjust the path based on feedback, and finally selecting the safe and reliable optimal solution. This invention significantly improves the reliability, objectivity, and security of complex decision-making by uniformly quantifying the target, evidence, and risk, and establishing a risk monitoring and closed-loop feedback mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and reasoning technology, in particular to a hierarchical attention large model reasoning system and method for complex decision-making. BACKGROUND

[0002] With the development of artificial intelligence technology, large language models (LLMs) have shown great capabilities in processing and generating complex texts, providing a new technical path for auxiliary systems facing complex decision-making. In high-risk fields such as enterprise project planning and financial investment, how to use the capabilities of large models to generate both creative and safe and reliable action plans has become an important research topic.

[0003] In existing applications, an end-to-end approach is usually adopted, where complex decision-making problems are directly input into large language models in the hope of generating complete action plans or strategy suggestions at one time. This method can provide novel ideas to some extent.

[0004] However, the inventors have found that this method has inherent defects. Since the reasoning process of large models is not transparent, the paths generated by it are often disconnected from objective reality, and there is a lack of quantifiable evaluation criteria to compare the pros and cons of different paths. It is difficult to implement a hierarchical attention mechanism, and it is difficult to effectively align and risk control multi-dimensional decision-making objectives, resulting in insufficient reliability of the final output and difficulty in being directly adopted in key decision-making.

[0005] Therefore, the present application proposes a hierarchical attention large model reasoning system and method for complex decision-making to solve the deficiencies of the prior art. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a hierarchical attention large model reasoning system and method for complex decision-making, which solves the problem that the large model reasoning mechanism lacks quantitative evaluation means for the alignment degree between decision-making objectives and reasoning paths when processing complex decision-making queries, and lacks a dynamic risk monitoring and feedback adjustment mechanism, resulting in insufficient efficiency and reliability of the reasoning process.

[0007] To achieve the above purpose, the present application is implemented by the following technical solutions:

[0008] The present application provides a hierarchical attention large model reasoning system for complex decision-making, which comprises:

[0009] a target intention module for receiving a complex decision-making query, decomposing the complex decision-making query into a target core, and constructing an ideal potential profile of a given dimension for the target core;

[0010] a path planning module configured to generate a plurality of reasoning paths in parallel according to the target cores and the ideal potential profile;

[0011] an operation verification module configured to retrieve evidence segments from a preset knowledge base through a to-be-verified path in the reasoning paths, and vectorize the evidence segments into evidence potential vectors of a predetermined dimension;

[0012] a potential processing module configured to aggregate the evidence potential vectors corresponding to the to-be-verified path to generate a polarized potential vector of the to-be-verified path;

[0013] a reasoning decision module configured to calculate a potential alignment degree of the to-be-verified path according to the polarized potential vector and the ideal potential profile, and monitor whether a dimension value of the polarized potential vector exceeds a preset potential safety threshold;

[0014] When the dimension value exceeds the potential safety threshold, the reasoning decision module generates a hedging subtask and feeds back to the path planning module to adjust the reasoning paths.

[0015] Preferably, the target intention module is specifically configured to:

[0016] receive a complex decision query input in the form of unstructured text through a natural language processing interface;

[0017] and perform deep semantic analysis and intention recognition on the complex decision query based on a preset large language model to extract one or more target cores representing core decisions;

[0018] and construct a potential coordinate system with the one or more target cores as the predetermined dimensions, then set ideal target values for each dimension in the potential coordinate system according to optimal targets, and finally combine the ideal target values of all dimensions to construct an ideal potential profile of the predetermined dimensions.

[0019] Preferably, the path planning module is specifically configured to:

[0020] use the target cores as initial planning conditions and use the ideal potential profile as a convergence target for path generation;

[0021] call a preset generative path planning model and perform a plurality of independent reasoning calculations based on the initial planning conditions and the convergence target to generate a plurality of reasoning paths, each independent reasoning calculation being performed by using different decoding parameters;

[0022] Each generated reasoning path is defined as a to-be-verified path.

[0023] Preferably, the operation verification module is specifically configured to:

[0024] parsing the to-be-verified path to extract core semantic keywords in each action step constituting the to-be-verified path;

[0025] performing a retrieval operation in the preset knowledge base with the core semantic keywords as a retrieval query to obtain one or more evidence segments most relevant to the action step;

[0026] for each of the obtained evidence segments, invoking a preset text analysis model to measure the support or opposition degree respectively embodied by the evidence segment in the given dimension, and quantifying the measurement result into a set of numerical values;

[0027] and combining the set of numerical values into a feature vector to generate an evidence potential energy vector corresponding to the evidence segment in the given dimension.

[0028] Preferably, the potential energy processing module is specifically configured to:

[0029] identifying and collecting all the evidence potential energy vectors corresponding to the current to-be-verified path to form a to-be-aggregated vector set;

[0030] performing a component-wise vector summation operation on all vectors in the to-be-aggregated vector set, that is, respectively performing algebraic addition on component values of all vectors in the same dimension, thereby calculating a single resultant vector;

[0031] taking the generated single resultant vector as a final aggregation result, and defining the aggregation result as a polarized potential energy vector of the to-be-verified path in the given dimension.

[0032] Preferably, the reasoning decision module is specifically configured to:

[0033] generating the potential energy alignment degree of the to-be-verified path by calculating the cosine similarity between the polarized potential energy vector and the ideal potential energy profile;

[0034] The calculation formula of the potential energy alignment degree is as follows:

[0035] ;

[0036] In the formula, is the potential energy alignment degree; is the polarized potential energy vector; is the ideal potential energy profile; denotes vector dot product operation.

[0037] Preferably, in the reasoning and decision-making module, the step of monitoring whether the dimension value of the polarization potential energy vector exceeds a preset potential energy safety threshold includes:

[0038] The potential energy safety threshold is defined as including A threshold vector of limit values, where The number of the given dimensions;

[0039] Each component of the polarization potential energy vector is checked individually. If at least one component satisfies the following condition, the dimension value is determined to exceed the potential energy safety threshold:

[0040] ;

[0041] in, This is the current dimension number being checked; The polarization potential vector is the first... The absolute value of the component values ​​in each dimension; The potential energy safety threshold is the first Preset limits in each dimension.

[0042] Preferably, the reasoning decision module is further configured to filter out reasoning paths that meet preset conditions based on the potential energy alignment and monitoring results in order to generate the final reasoning result;

[0043] The final reasoning result includes:

[0044] The optimal reasoning path is selected from the reasoning paths that meet the preset conditions, based on the highest potential energy alignment score.

[0045] The potential alignment score corresponding to the optimal inference path;

[0046] A verification state in which all dimensions of the polarization potential vector that confirms the optimal inference path do not exceed the potential safety threshold.

[0047] A second aspect of the present invention provides a hierarchical attention-based large model inference method for complex decision-making, the method comprising the following steps:

[0048] Receive complex decision queries, decompose the complex decision queries into target kernels, and construct an ideal potential energy profile of a predetermined dimension for the target kernels;

[0049] Based on the target kernel and the ideal potential energy profile, multiple parallel inference paths are generated;

[0050] Evidence fragments are retrieved from a preset knowledge base through the path to be verified in the reasoning path, and the evidence fragments are vectorized into evidence potential vectors of a predetermined dimension.

[0051] Aggregate the evidence potential vectors corresponding to the path to be verified to generate the polarization potential vector of the path to be verified.

[0052] The potential energy alignment of the path to be verified is calculated based on the polarization potential energy vector and the ideal potential energy profile, and the dimension value of the polarization potential energy vector is monitored to see if it exceeds the preset potential energy safety threshold.

[0053] When the dimensional value exceeds the potential energy safety threshold, a hedging subtask is generated, and the process returns to the step of generating multiple parallel inference paths for adjustment.

[0054] Preferably, the method further includes: based on the potential energy alignment and the monitoring results, selecting inference paths that meet preset conditions to generate a final inference result.

[0055] This invention provides a hierarchical attention-based large-scale model inference system and method for complex decision-making. It offers the following advantages:

[0056] 1. This invention transforms a qualitative problem concerning complex decision-making into a mathematical measurement problem within a multi-dimensional vector space by introducing concepts such as target kernel, ideal potential energy profile, and polarized potential energy vector. By calculating the cosine similarity between the polarized potential energy vector and the ideal potential energy profile, the system can accurately quantify the degree of consistency between any reasoning path and the final decision objective. This provides a reliable quantitative evaluation method for the entire reasoning process, solves the problem of the lack of alignment-based quantitative evaluation methods in existing technologies, and significantly improves the objectivity and interpretability of the decision-making process.

[0057] 2. This invention achieves pre-identification of potential risks in the decision-making path by pre-setting multi-dimensional potential energy safety thresholds and utilizing a hierarchical attention mechanism to monitor the polarization potential energy vector of each path to be verified in real time. Once the potential energy value in any dimension exceeds the safety limit, the large model system can generate a hedging subtask and trigger dynamic adjustments to the inference path. This closed-loop feedback mechanism solves the problem of the lack of dynamic risk monitoring and feedback adjustment mechanisms in existing technologies, effectively avoiding potential negative outcomes and significantly enhancing the safety of the final decision.

[0058] 3. This invention significantly improves the reasoning efficiency for complex decisions by generating multiple inference paths in parallel and using a unified potential energy framework for rapid evaluation and selection. Simultaneously, by combining quantitative ranking of potential energy alignment with a risk monitoring mechanism, it ensures that the ultimately selected optimal inference path is not only highly aligned with the decision objective but also safe at the execution level. This solution integrates evaluation, monitoring, and closed-loop control, solving the problems of low inference efficiency and insufficient reliability in existing technologies, and ensuring the overall quality of the final decision result. Attached Figure Description

[0059] Figure 1 This is a diagram of a hierarchical attention-based large-scale model inference system architecture for complex decision-making according to the present invention.

[0060] Figure 2 This is a flowchart of a hierarchical attention-based large model inference method for complex decision-making according to the present invention;

[0061] Figure 3 This is a flowchart of the path generation and quantization process of the present invention;

[0062] Figure 4 This is a flowchart illustrating the path aggregation and decision-making process of the present invention.

[0063] Among them, 100 is the target intent module; 200 is the path planning module; 300 is the operation verification module; 400 is the potential energy processing module; and 500 is the reasoning and decision-making module. Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] See attached document Figure 1 , Figure 1 This is an architecture diagram of a hierarchical attention-based large-scale model inference system for complex decision-making according to an embodiment of the present invention. The hierarchical attention-based large-scale model inference system for complex decision-making provided by the present invention may include: a target intent module 100, a path planning module 200, an operation verification module 300, a potential energy processing module 400, and an inference decision-making module 500.

[0066] The target intent module 100 receives a complex decision query, decomposes the query into one or more target kernels, and constructs an ideal potential energy profile of a predetermined dimension based on these kernels. This ideal potential energy profile provides a quantitative benchmark target for subsequent path evaluation.

[0067] The path planning module 200, connected to the target intent module 100, is used to generate multiple parallel, distinct inference paths based on the target kernel and the ideal potential energy profile. Each generated inference path is defined as a path to be verified.

[0068] The operation verification module 300, which is connected to the path planning module 200, is used to receive a path to be verified and retrieve relevant evidence fragments from a preset knowledge base according to the action steps in the path. Then, each evidence fragment is converted into an evidence potential vector of a predetermined dimension.

[0069] The potential energy processing module 400, which is connected to the operation verification module 300, is used to aggregate all evidence potential energy vectors related to the current path to be verified and generate a single polarization potential energy vector that can characterize the overall effect of the path.

[0070] The reasoning and decision-making module 500 is connected to the target intent module 100 and the potential energy processing module 400, respectively. This module calculates the potential energy alignment of the path to be verified based on the polarized potential energy vector and the ideal potential energy profile. Simultaneously, this module monitors whether each dimension component of the polarized potential energy vector exceeds a preset potential energy safety threshold.

[0071] If the dimensional component exceeds the safety threshold, the inference decision module 500 will generate a hedging subtask and feed it back to the path planning module 200 to adjust the path. If it does not exceed the threshold, the final inference result will be generated based on the potential energy alignment score.

[0072] See attached document Figure 1 and Figure 2 , Figure 2 This is a flowchart of path generation and quantization according to an embodiment of the present invention. The target intent module 100 is the entry point of the entire inference system, which is responsible for transforming an unstructured initial query oriented towards complex decision-making into a structured benchmark target that can be used for subsequent quantization calculations. In a specific embodiment, the implementation of this module includes the following processing steps.

[0073] First, the target intent module 100 receives complex decision queries input in unstructured text form through a natural language processing interface. The natural language processing interface is a software interface, such as an application programming interface (API) endpoint, that defines the protocol and format for receiving UTF-8 encoded text data streams.

[0074] Upon receiving a query, the target intent module 100 invokes a pre-defined large model (large language model) for processing. This pre-defined large language model is a deep neural network model based on the Transformer architecture. This deep neural network model has been pre-trained on a general text corpus containing hundreds of millions or billions of documents, enabling it to perform deep semantic analysis and intent recognition on long texts. The model parses the input unstructured text to identify and extract one or more target kernels representing core decision-making objectives. For example, for a query assessing the feasibility and risks of a company entering the European new energy market within the next two years, the extracted target kernels could be market share growth, brand image enhancement, financial investment risk, and supply chain stability.

[0075] After extracting the target kernel, the target intent module 100 uses one or more target kernels as predetermined dimensions to construct a multi-dimensional potential energy coordinate system, the number of dimensions being... This is equal to the number of target kernels extracted. Establishing this coordinate system is the first step in quantifying the decision-making process, providing a unified mathematical space for all subsequent evaluations.

[0076] Subsequently, within this potential energy coordinate system, the target intent module 100 assigns an ideal target value to each dimension, i.e., each target kernel, based on preset business rules or optimal goals. For example, market share growth and brand image enhancement, as desired goals, can be assigned positive values ​​(e.g., +10); while financial investment risk and supply chain stability, as risks that need to be avoided or controlled, can be assigned negative values ​​(e.g., -10). The absolute magnitude of these values ​​characterizes the importance of that dimension.

[0077] Finally, the target intent module 100 combines the ideal target values ​​across all dimensions to form a... A three-dimensional vector, defined as the ideal potential energy profile, transforms an abstract decision objective into a precise mathematical object, thus addressing the current lack of quantitative evaluation methods for decision objectives. This profile provides the highest-level focus for subsequent hierarchical attention mechanisms.

[0078] See attached document Figure 1 and Figure 2 The path planning module 200's input is connected to the output of the target intent module 100. Within the reasoning framework of the larger model, this module is responsible for generating specific and executable action plans for the decision objectives determined by the upper-level modules. In a specific embodiment, the path planning module 200 performs the following processing steps.

[0079] First, the path planning module 200 receives the target kernel and ideal potential profile output by the target intent module 100. In this process, the target kernel is used as the initial planning condition for text generation, for example, combined into an initial prompt input to the generative model. The ideal potential profile, on the other hand, is used as the macroscopic convergence target for path generation, providing directional guidance for the entire generation process and ensuring that the generated path remains consistent with the final target at the target level.

[0080] The path planning module 200 calls a preset generative path planning model. The path planning module 200 is a deep neural network model based on the Transformer architecture that has been fine-tuned on specific datasets such as planning and task decomposition. This model receives the initial planning conditions and performs multiple independent inference calculations to generate multiple parallel inference paths.

[0081] A reasoning path is a structured text output by a generative path planning model based on the initial planning conditions. Formally, this text represents a coherent action plan or analytical outline consisting of multiple logical steps. Each step describes a specific operation or consideration recommended to achieve the final goal. For example, for a complex market entry decision, a reasoning path could be a text containing multiple numbered steps, such as:

[0082] Conduct due diligence on the laws and regulations of the target market;

[0083] Analyze the pricing strategies and market share of the top three competitors;

[0084] Develop a three-phase marketing and promotion plan and calculate the budget.

[0085] Therefore, each reasoning path represents a complete and independent solution proposed by the system for the initial query.

[0086] To ensure the content diversity of the generated multiple inference paths, each independent inference calculation is performed using distinct decoding parameters. These decoding parameters control the neural network model's strategy for selecting the next word when generating text. In one embodiment, path diversification is achieved by adjusting one or more of the following parameters:

[0087] Randomness Factor: This parameter controls the degree of randomness in the generated results. A higher factor value (e.g., 0.9) encourages the model to explore more possibilities when choosing the next word, resulting in more diverse and creative paths; while a lower factor value (e.g., 0.2) encourages the model to choose the word with the highest probability, resulting in more logically deterministic and conservative paths.

[0088] Sampling parameter: This parameter limits the selection of the model to those with the highest probability. Within the candidate words. Change The value can adjust the breadth of the generated paths.

[0089] (Nucleus) Sampling Parameter: This parameter is derived from a cumulative probability exceeding a threshold. Sampling is performed from the minimum candidate word set. Change The value can balance the uniqueness and coherence of the path.

[0090] By setting different combinations of decoding parameters for each inference calculation, the system can explore the solution space and generate a set of inference paths that differ in content and strategy. This is a key step in the entire hierarchical attention mechanism, involving divergent thinking at the path planning level, and provides alternative solutions for subsequent evaluation and convergence. Each generated inference path is defined as a path to be verified and is sent to the operation verification module 300 for further processing.

[0091] See attached document Figure 1 and Figure 2 The input of the operation verification module 300 is connected to the output of the path planning module 200. This module is a key link in realizing the hierarchical attention mechanism, from reasoning path to specific evidence. Its core function is to transform a path to be verified generated by a large model into a quantifiable mathematical object by comparing it with external knowledge. In a specific embodiment, the operation verification module 300 performs the following processing steps.

[0092] First, the operation verification module 300 receives a path to be verified from the path planning module 200 and parses the path to extract the core semantic keywords in each action step that constitutes the path. This parsing process is accomplished by calling a preset language model. In one embodiment, the preset language model is a deep neural network model based on a Transformer encoder architecture. This model is pre-trained on a general corpus and then fine-tuned on a dataset containing named entity annotations and keyword annotations. Therefore, the model can identify and separate key noun phrases and verb phrases of preset categories from the input action step text as the core semantic keywords.

[0093] Secondly, this module uses the extracted core semantic keywords as the retrieval query to perform a retrieval operation in a preset knowledge base to obtain one or more evidence fragments optimally related to the action step. The preset knowledge base is a vector database, which pre-stores a large number of document records related to the decision-making domain, and each record has been encoded into a high-dimensional vector by a text embedding model. The operation verification module 300 calls the same text embedding model to convert the query statement composed of the core semantic keywords into a query vector, and then executes the query in the vector database. The nearest neighbor (k-NN) algorithm is used to retrieve one or more pieces of evidence in the vector space that have the highest cosine similarity to the query vector.

[0094] Next, for each acquired piece of evidence, the operation verification module 300 invokes a preset text analysis model to measure the degree of support or opposition reflected by the evidence piece on the predetermined dimensions, and quantifies this measurement result into a set of numerical values. The text analysis model is a deep neural network based on the Transformer architecture, which is designed for a given set of... The regression task of the dimension was trained, in which The number of dimensions defined by the target intent module 100. The model receives a text fragment of evidence as input and outputs a... An array of floating-point numbers, i.e., the set of values. The nth floating-point number in the array. The numerical value represents the significance of the evidence fragment for the first... The degree of support (positive value) or opposition (negative value) for a target kernel is represented by its absolute value, which indicates the strength of that support or opposition.

[0095] Finally, the operation verification module 300 combines the set of values ​​into a feature vector to generate an evidence potential vector of a predetermined dimension corresponding to the evidence fragment. This combination process involves taking the output from the previous step, which includes... An array of floating-point numbers can be directly constructed into a... A dimensional vector. This vector is defined as the evidence potential vector. Through the above steps, this module associates each step in an abstract reasoning path with real-world evidence and fully quantifies this association. This process is the foundation for the entire reasoning system to reliably evaluate complex decisions, solving the problem of existing technologies lacking factual basis and quantitative evaluation methods for generated paths.

[0096] See attached document Figure 1 and Figure 3 , Figure 3This is a path aggregation and decision-making flowchart according to an embodiment of the present invention. The input end of the potential energy processing module 400 is connected to the output end of the operation verification module 300. The potential energy processing module 400 undertakes the function of information aggregation in the entire inference system and is a key step in realizing the hierarchical attention mechanism by aggregating discrete evidence points into a holistic path view. In a specific embodiment, the potential energy processing module 400 performs the following processing steps.

[0097] First, the potential energy processing module 400 identifies and aggregates all the evidence potential energy vectors corresponding to the current path to be verified, forming a vector set to be aggregated. Specifically, each evidence potential energy vector output from the operation verification module 300 is accompanied by a unique identifier associated with its respective path to be verified. The potential energy processing module 400 groups the incoming vector stream according to this unique identifier, grouping all evidence potential energy vectors with the same identifier into the same processing queue. This queue can be a vector array or a list in terms of data structure, i.e., the vector set to be aggregated.

[0098] Secondly, this module performs a component-wise vector summation operation on all vectors in the vector set to be aggregated, that is, algebraically adding the component values ​​of all vectors in the same dimension to calculate a single composite vector. The algorithm steps of this calculation process are as follows:

[0099] Initialize a zero vector of dimension ,in The number of dimensions is given.

[0100] Iterate through each evidence potential vector in the set of vectors to be aggregated. (in From 1 to , (This represents the total number of vectors in the set).

[0101] For each vector Then iterate through its range from 1 to... Each dimension component .

[0102] Components The values ​​are accumulated into the zero vector The corresponding number On each component.

[0103] After traversing all After processing each vector, the final composite vector is obtained. .

[0104] The mathematical expression of this process is defined by the following formula:

[0105] ;

[0106] In the formula, the composite vector The Each component (for The calculation method is as follows:

[0107] ;

[0108] In the formula, It is the first The first evidence potential vector Each dimension has a component. This summation operation can be computationally accelerated in parallel using the Single Instruction Multiple Data (SIMD) instruction set of the Central Processing Unit (CPU) or the Graphics Processing Unit (GPU).

[0109] Finally, the potential energy processing module 400 uses the generated single composite vector as the final aggregation result and defines the aggregation result as the polarized potential energy vector of the predetermined dimension of the path to be verified. Here, polarization means that, through algebraic summation, multiple positive values ​​representing support and multiple negative values ​​representing opposition cancel each other out in each dimension, ultimately causing the component values ​​of the composite vector in that dimension to tend towards the polarity (positive or negative) of the net effect.

[0110] This aggregation operation unifies the quantified effects from different pieces of evidence, which may support or contradict each other, across each objective dimension to arrive at a net effect value. This process addresses the difficulty of comprehensively evaluating multiple pieces of evidence containing contradictory information in existing large models when facing complex decision-making. By generating a single polarization potential vector, this module provides a clear, complete, and quantifiable evaluation object for the subsequent inference and decision-making module 500.

[0111] See attached document Figure 1 and Figure 3 The input terminals of the inference decision module 500 are connected to the output terminals of the target intent module 100 and the potential energy processing module 400, respectively. This module is the final decision-making unit of the entire hierarchical attention large model inference system, responsible for comparing, filtering, and making a final decision on all quantified parallel inference paths in response to the initial complex decision query. In a specific embodiment, the inference decision module 500, as the endpoint of the data processing pipeline, performs the following processing steps on each path to be verified:

[0112] First, for each path to be verified generated by the path planning module 200 and processed by subsequent modules, the inference and decision-making module 500 receives two core parameters. 3D vector data object: The polarization potential energy vector (denoted as ) calculated by the potential energy processing module 400, representing the overall effect of the path. ), and the ideal potential energy profile representing the global final goal (denoted as ) generated by the target intent module 100. ).

[0113] The inference and decision-making module 500 generates the potential energy alignment of the path to be verified by calculating the cosine similarity between the polarization potential energy vector and the ideal potential energy profile. The purpose of this calculation is to... In a 3D potential energy coordinate system, this is the angle between the direction of the actual effect of the path and the direction of the ideal target. The closer this value is to 1, the more closely the path matches the target at the target level; conversely, a smaller value indicates deviation or even opposition. This calculation is performed using standard floating-point arithmetic, and the formula for calculating the potential energy alignment is as follows:

[0114] ;

[0115] In the formula, The potential energy alignment degree; Let be the polarization potential energy vector; This is the ideal potential energy profile; This represents the vector dot product operation.

[0116] Simultaneously, the inference and decision-making module 500 performs a parallel monitoring task: monitoring whether the dimension value of the polarization potential energy vector exceeds a preset potential energy safety threshold. The purpose of this step is to prevent the system from adopting a path with unacceptable risk in a critical dimension. This monitoring step specifically includes:

[0117] The potential energy safety threshold is defined as including A threshold vector of limits (denoted as ) ),in The number of the given dimensions, and Each component in All are positive numbers. Subsequently, the polarization potential vector... Each dimension component is checked one by one. When at least one dimension component satisfies the following condition, it is determined that the dimension value exceeds the potential energy safety threshold:

[0118] ;

[0119] in, The dimension number currently being checked (from 1 to ...) ); The polarization potential vector is the first... The absolute value of the component values ​​in each dimension; The potential energy safety threshold is the first Preset limits in each dimension.

[0120] When the dimension value exceeds the potential energy safety threshold, the path is marked as unsafe. At this time, the inference decision module 500 generates a hedging subtask and feeds it back to the path planning module 200 to adjust the content of the inference path. Specifically, the inference decision module 500 addresses the exceeded dimension (i.e., the first...) (Each dimension) generates a text instruction aimed at reducing the risk of that dimension as the hedging subtask. For example, if the value of the financial investment risk dimension exceeds a threshold, a text instruction in the form of "Please supplement specific measures to reduce financial investment risk" can be generated. After receiving the hedging subtask, the path planning module 200 uses it as a new generation constraint to adjust and regenerate the content of the inference path. This closed-loop feedback mechanism is an important beneficial effect of the present invention, enabling the system to dynamically self-correct rather than simply abandoning risky paths.

[0121] Finally, the inference decision module is also used to filter out inference paths that meet preset conditions based on the potential energy alignment and monitoring results to generate the final inference result. The filtering logic is as follows:

[0122] First, exclude all paths marked as unsafe. Then, among all safe paths, sort them in descending order based on their potential alignment score.

[0123] The final reasoning result is a structured data object, which includes:

[0124] Optimal inference path: A textual description of the optimal inference path selected from the inference paths that meet the preset conditions, based on the highest potential alignment score.

[0125] Potential alignment score: The specific numerical value calculated by the above formula corresponding to the optimal inference path.

[0126] Verification status: A Boolean or enumerated state that confirms that all dimension values ​​of the polarization potential energy vector of the optimal inference path do not exceed the potential energy safety threshold.

[0127] This dual evaluation mechanism, which combines orientation alignment and risk threshold checks, ensures that the system outputs not only the theoretically optimal solution but also a safe and feasible solution in practice, greatly improving the reliability and practicality of the large model in complex real-world decision-making scenarios.

[0128] See attached document Figure 2 , Figure 2This is a flowchart illustrating a hierarchical attention-based large-scale model inference method for complex decision-making according to an embodiment of the present invention. The hierarchical attention-based large-scale model inference method for complex decision-making is executed on one or more server devices, completing an end-to-end decision-making process by invoking different neural network models. The method includes the following steps:

[0129] S1. Receive complex decision queries, decompose the complex decision queries into target kernels, and construct an ideal potential energy profile of a predetermined dimension for the target kernels;

[0130] This step is executed by the target intent module 100. First, the system receives an unstructured, complex decision query through a natural language processing interface. Then, a pre-defined large model performs deep semantic analysis on the query text, extracting one or more target kernels representing the core target. These target kernels are defined as orthogonal, predetermined dimensions, and are assigned ideal target values ​​representing the desired direction and importance according to pre-defined rules, thus constructing a target kernel. The ideal potential energy profile vector of dimension. This step sets the highest-level, quantifiable convergence objective for the entire hierarchical attention mechanism.

[0131] S2. Based on the target kernel and the ideal potential energy profile, generate multiple parallel inference paths;

[0132] This step is performed by the path planning module 200. The system uses the target kernel extracted in S1 as the initial generation condition and calls a generative large model to perform multiple independent inference calculations. Different decoding parameters (such as temperature, etc.) are set for each calculation. (Through sampling, etc.), the model can generate a set of paths to be validated that differ significantly in strategy and content. This step constitutes a stage of divergent thinking, providing a wealth of alternatives for subsequent evaluation.

[0133] S3. Retrieve evidence fragments from a preset knowledge base through the path to be verified in the reasoning path, and vectorize the evidence fragments into an evidence potential vector of a predetermined dimension;

[0134] This step is executed by the operation verification module 300. For each path to be verified generated in S2, the system first parses its action steps to extract core keywords. Then, using these keywords as queries, it retrieves the most semantically relevant evidence fragments from a preset vector knowledge base. Finally, the system calls a preset, targeted... The text analysis model, trained on a dimensional regression task, analyzes each piece of evidence and quantifies it into a single... A dimensional evidence potential vector, where each component of the vector represents the degree to which the evidence supports or opposes the corresponding target dimension.

[0135] S4. Aggregate the evidence potential vectors corresponding to the path to be verified to generate the polarization potential vector of the path to be verified.

[0136] This step is performed by the potential energy processing module 400. The system aggregates all evidence potential energy vectors generated in S3 that are related to the same path to be verified. This aggregation operation is specifically a component-wise vector summation, algebraically superimposing the influence of all evidence in each dimension to calculate a single polarization potential energy vector that can represent the overall effect of the path.

[0137] S5. Calculate the potential energy alignment of the path to be verified based on the polarization potential energy vector and the ideal potential energy profile, and monitor whether the dimension value of the polarization potential energy vector exceeds the preset potential energy safety threshold.

[0138] This step is performed by the inference and decision-making module 500. The system performs a dual evaluation on the polarization potential energy vector generated by S4. On the one hand, the potential energy alignment is obtained by calculating the cosine similarity between it and the ideal potential energy profile generated by S1 to assess the correctness of the direction. On the other hand, the absolute value of each component is compared with a preset potential energy safety threshold vector to monitor for the existence of unacceptable single-point risks.

[0139] S6. When the dimension value exceeds the potential energy safety threshold, generate a hedging subtask and return to the step of generating multiple parallel inference paths for adjustment.

[0140] This step is a closed-loop feedback loop in the decision-making process, completed collaboratively by the inference decision-making module 500 and the path planning module 200. If a value in a certain dimension of the polarization potential energy vector exceeds a safety threshold in S5, the inference decision-making module generates a text instruction aimed at reducing the risk in that dimension as a hedging subtask and feeds it back to the path planning module. The path planning module then performs a new inference generation, using this hedging subtask as an additional constraint to generate a modified, lower-risk inference path.

[0141] S7. When the dimension value does not exceed the potential energy safety threshold, based on the potential energy alignment degree and the monitoring result, a reasoning path that meets the preset conditions is selected to generate the final reasoning result.

[0142] This step is the final decision output stage, executed by the inference decision module 500. For all paths that pass S5 security monitoring, the system sorts them in descending order according to their potential alignment scores and selects the path with the highest score as the optimal inference path. Finally, the system outputs a structured result containing the optimal path text, its alignment score, and security verification status, completing the entire inference process for complex decisions.

[0143] The following explanation uses a hypothetical failure scenario as an example.

[0144] Application scenario: Assessing the feasibility of a technology company entering the European new energy market.

[0145] Step 1: Input query and target setting (Target Intent Module 100)

[0146] Query received: The system received a complex decision query from the user: Please assess the feasibility, opportunities and risks of our company (a consumer electronics company) entering the European electric vehicle charging station market within the next two years.

[0147] Decompose the target kernels: The system calls the large model and decomposes the query into four core target kernels: [market share growth (opportunity), brand image enhancement (opportunity), supply chain risk (risk), regulatory compliance cost (risk)].

[0148] Constructing the Ideal Profile: Based on preset rules, the system sets ideal target values ​​for these four target cores (dimensions), forming an ideal potential energy profile. The ideal state represented is: significantly increasing market share and brand image, while keeping supply chain risks and regulatory costs at extremely low levels.

[0149] Step 2: Generate diverse action plans (Route planning module 200)

[0150] Based on the aforementioned target kernel, the system generates multiple inference paths with varying content in parallel by adjusting internal randomness factors and other parameters. For example:

[0151] Reasoning Path A (Active Expansion):

[0152] Acquire a mid-sized European charging pile manufacturer.

[0153] Quickly integrate our existing power management technologies into their product line.

[0154] Conduct large-scale marketing campaigns in key countries.

[0155] Reasoning Path B (Cautious Pilot):

[0156] First, establish wholly-owned subsidiaries in countries D and N.

[0157] We partnered with a few local car dealerships to conduct small-scale pilot installations.

[0158] Based on market feedback data, assess the next expansion plan.

[0159] Step 3: Verify the path and quantify the impact (Operational verification module 300)

[0160] The system verifies every step of every path. Take the second step of inference path B, "Cooperate with local distributors to conduct small-scale pilot installations," as an example:

[0161] Evidence retrieval: The system retrieves relevant evidence from its internal knowledge base (which includes industry reports, news, and regulatory documents).

[0162] Fragments of evidence found:

[0163] Evidence 1 (positive): "A report from the D Country Automobile Association shows that dealers are open to bundling high-quality charging stations, believing it can enhance customer value."

[0164] Evidence 2 (Negative): "Industry analysis points out that the charging pile market in Country N is highly competitive, and the channel construction costs for new brands entering the market are extremely high."

[0165] Quantitative Evidence: The system transforms each piece of evidence into an evidence potential vector. Evidence 1 will generate a positive value in the market share growth dimension; Evidence 2 will generate a significant negative value in the supply chain risk dimension (reflected here as channel costs).

[0166] Step 4: Overview of the aggregation effect formation path (potential energy processing module 400)

[0167] The system algebraically sums all the evidence potential vectors generated by all steps of reasoning path B to generate a single polarization potential vector. This vector comprehensively reflects the net effect of path B: it has some potential for market share growth, but exhibits higher risk in the supply chain risk dimension.

[0168] Step 5: Decision-making, correction, and output (Reasoning and decision-making module 500)

[0169] Assessment and monitoring:

[0170] The system calculated that the potential alignment score of inference path A was very high, but its "acquisition" operation triggered the potential safety threshold of the financial investment risk dimension, so the path was marked as unsafe.

[0171] When a path is marked as unsafe, the system does not simply discard it. Instead, the inference decision module 500 generates a hedging subtask and feeds it back to the path planning module 200 to adjust and optimize the inference path. For example, the system might generate a specific correction instruction: "High financial risk detected. Please modify path A, replacing the 'acquisition' option with the option of 'establishing cooperation or a joint venture with the target company,' and reassess."

[0172] Meanwhile, the system continues to evaluate other paths. Inference path B has a moderate potential alignment score, and all component values ​​of the risk dimensions do not exceed the safety threshold, so it is marked as safe.

[0173] Filtering and Output:

[0174] In this decision loop, the system first eliminates the option of immediately adopting the unsafe path A, and selects the only currently safe path B as the optimal choice. Finally, a structured result is output to the user:

[0175] Optimal reasoning path: A cautious pilot approach is recommended. First, conduct small-scale pilot programs in countries D and N, then determine subsequent steps based on market feedback.

[0176] Potential energy alignment: 0.78 (indicating a good match with the ideal target).

[0177] Verification status: Safe. All key risk dimensions are within controllable thresholds.

[0178] Key reason: This approach can effectively control the risks of initial investment and channel development while achieving market growth, and avoid the financial risks that may result from large-scale mergers and acquisitions.

[0179] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hierarchical attention large model reasoning system for complex decision-making, characterized in that, The method comprises the following steps: a target intention module is configured to receive a complex decision query, decompose the complex decision query into target cores, and construct an ideal potential profile of a predetermined dimension for the target cores; a path planning module is configured to generate a plurality of parallel reasoning paths according to the target cores and the ideal potential profile; an operation verification module is configured to retrieve evidence fragments from a preset knowledge base through a to-be-verified path in the reasoning paths, and vectorize the evidence fragments into an evidence potential vector of a predetermined dimension; a potential processing module is configured to aggregate the evidence potential vectors corresponding to the to-be-verified path to generate a polarized potential vector of the to-be-verified path; a reasoning decision module is configured to calculate a potential alignment degree of the to-be-verified path according to the polarized potential vector and the ideal potential profile, and monitor whether a dimension value of the polarized potential vector exceeds a preset potential safety threshold; when the dimension value exceeds the potential safety threshold, the reasoning decision module generates a hedging sub-task and feeds it back to the path planning module to adjust the reasoning paths; the target intention module is specifically configured to: receive a complex decision query input in the form of unstructured text through a natural language processing interface; perform deep semantic analysis and intention recognition on the complex decision query based on a preset large language model to extract one or more target cores representing core decisions; construct a potential coordinate system with the one or more target cores as the predetermined dimensions, then set ideal target values for each dimension in the potential coordinate system according to optimal targets, and finally combine the ideal target values of all dimensions to construct an ideal potential profile of the predetermined dimensions; the operation verification module is specifically configured to: parse the to-be-verified path to extract core semantic keywords in each action step constituting the to-be-verified path; use the core semantic keywords as a retrieval query to perform a retrieval operation in the preset knowledge base to obtain one or more evidence fragments most relevant to the action steps; for each obtained evidence fragment, call a preset text analysis model to measure the degree of support or opposition embodied by the evidence fragment in the predetermined dimensions, and quantify the measurement result as a set of numerical values; combine the set of numerical values into a feature vector to generate an evidence potential vector of the predetermined dimensions corresponding to the evidence fragment.

2. The hierarchical attention large model reasoning system for complex decision-making according to claim 1, wherein, the path planning module is specifically configured to: use the target cores as initial planning conditions and use the ideal potential profile as a convergence target for path generation; call a preset generative path planning model and perform multiple independent reasoning calculations based on the initial planning conditions and the convergence target to generate a plurality of reasoning paths, each independent reasoning calculation being performed by using different decoding parameters; wherein each generated reasoning path is defined as a to-be-verified path.

3. The hierarchical attention large model reasoning system for complex decision-making of claim 1, wherein, the potential processing module is specifically configured to: identify and collect all evidence potential vectors corresponding to the current to-be-verified path to form a to-be-aggregated vector set; performing a component-wise vector sum operation on all vectors in the set of vectors to be aggregated, i.e., algebraically adding component values of all vectors in the same dimension respectively, thereby calculating a single resultant vector; taking the generated single resultant vector as a final aggregation result, and defining the aggregation result as the polarization potential energy vector of the given dimension of the path to be verified.

4. The hierarchical attention large model reasoning system for complex decision-making of claim 1, wherein, The reasoning decision module is specifically configured to: generate the potential energy alignment degree of the path to be verified by calculating the cosine similarity between the polarization potential energy vector and the ideal potential energy profile; The formula for calculating the potential energy alignment degree is as follows: ； wherein is the potential alignment degree; is the polarization potential vector; is the ideal potential profile; denotes the vector dot product operation.

5. The hierarchical attention large model reasoning system for complex decision-making of claim 1, wherein, In the reasoning decision module, the step of monitoring whether the dimension value of the polarization potential energy vector exceeds a preset potential energy safety threshold comprises: defining the potential energy safety threshold as a threshold vector comprising limit values, wherein is the number of the given dimensions; and each dimension component of the polarization potential energy vector is checked one by one, and when at least one dimension component satisfies the following condition, it is determined that the dimension value exceeds the potential energy safety threshold: ； wherein, is the current dimension number of the verification; is the absolute value of the component value of the polarization potential vector in the dimension; is the preset limit value of the potential energy security threshold in the dimension.

6. The hierarchical attention large model reasoning system for complex decision-making of claim 1, wherein, The reasoning decision module is further configured to filter out reasoning paths that meet a preset condition based on the potential energy alignment degree and the monitoring result to generate a final reasoning result; The final reasoning result includes: the optimal reasoning path selected from the reasoning paths that meet the preset condition according to the highest potential energy alignment degree score; the potential energy alignment degree score corresponding to the optimal reasoning path; a verification state confirming that all dimension values of the polarization potential energy vector of the optimal reasoning path do not exceed the potential energy safety threshold.

7. A hierarchical attention large model reasoning method for complex decision-making, applied to a hierarchical attention large model reasoning system for complex decision-making according to any one of claims 1-6, characterized in that, The method comprises the following steps: receiving a complex decision query, decomposing the complex decision query into a target core, and constructing an ideal potential energy profile of a given dimension for the target core; generating a plurality of parallel reasoning paths according to the target core and the ideal potential energy profile; retrieve evidence fragments from a preset knowledge base through the path to be verified in the reasoning paths, and vectorize the evidence fragments into evidence potential energy vectors of a given dimension; aggregate the evidence potential energy vectors corresponding to the path to be verified to generate a polarization potential energy vector of the path to be verified; calculate the potential energy alignment degree of the path to be verified according to the polarization potential energy vector and the ideal potential energy profile, and monitor whether the dimension value of the polarization potential energy vector exceeds a preset potential energy safety threshold; when the dimension value exceeds the potential energy safety threshold, generate a hedging subtask, and return to execute the step of generating a plurality of parallel reasoning paths to adjust.

8. The hierarchical attention large model reasoning method for complex decision-making according to claim 7, characterized in that, Further comprising: filtering out reasoning paths that meet a preset condition based on the potential energy alignment degree and the monitoring result to generate a final reasoning result.

Images