Intelligent generation method for financial supervision report based on large model

Abstract

The application provides a large model-based financial supervision report intelligent generation method, relates to the technical field of data processing systems, and first constructs a supervision three-view structure containing regulatory provisions, supervision question and answer and penalty cases, and obtains the potential vector representation of the supervision provisions by using a large model and a graph embedding model. It can output the compliance risk score in the multi-explanation scene according to the institution state and the report result. On this basis, an initial report strategy in the form of an executable report graph is generated by the large model, a strict explanation and confrontation scene is constructed by introducing a supervision counterparty agent, the report strategy is optimized in the way of maximum-minimum game, and a strategy result robust to the most unfavorable supervision explanation is obtained. After the strategy converges, the supervision elasticity tensor is further calculated to depict the sensitivity of the report result to potential supervision changes, and the tensor is used to realize the self-evolution update of the newly added or revised provisions. The application can improve the automation, robustness and explainable consistency of the supervision report.

Description

Technical Field

[0001] This invention relates to the field of data processing system technology, specifically to a method for intelligent generation of financial regulatory reports based on large models. Background Technology

[0002] Financial regulatory reporting is a crucial step for commercial banks, securities firms, and insurance companies in fulfilling their regulatory obligations. It typically requires institutions to submit structured reports and explanatory materials to regulatory agencies on a regular basis, in accordance with the regulations, reporting guidelines, and regulatory interpretations issued by the regulatory authorities. Currently, most reporting logic relies on manual interpretation of regulations, manual formulation of interpretation rules, and manual maintenance of the reporting process. With frequent updates to regulatory policies, rapidly increasing regulatory complexity, and diversified reporting scenarios, manual methods are becoming increasingly difficult to handle inconsistencies in rules, cross-regulation relationships, and interpretation of abnormal scenarios, easily leading to errors, omissions, or inconsistent understandings of regulatory interpretations.

[0003] In recent years, regulatory authorities have emphasized the interpretability, verifiability, and consistency of reporting, requiring institutions to clearly explain their reporting logic and quickly reproduce the reporting process during inspections. However, traditional scripted reporting procedures struggle to cope with differing regulatory interpretations and cannot be quickly adjusted after regulatory clauses change. In regulatory inspections or penalty cases, many problems stem from institutions' insufficient understanding of the specific scope of application of clauses or their failure to maintain consistent reporting logic in extreme scenarios. Therefore, there is an urgent need for a technology capable of understanding large amounts of regulatory data, uniformly modeling regulatory interpretations, and automatically generating executable reporting strategies; it must also be robust under different regulatory interpretations and able to quickly adapt to newly added or revised clauses to reduce compliance risks and improve communication efficiency between regulators and institutions. Summary of the Invention

[0004] Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent generation method for financial regulatory reports based on large models, thus solving the problems of existing technologies.

[0006] Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: a financial regulatory reporting intelligent generation method based on a large model, comprising the following steps: Sp1 regulatory multi-source corpus modeling: acquiring regulatory corpus such as regulatory regulations texts, regulatory Q&A guidelines, and regulatory penalty cases, preprocessing them, and constructing a regulatory three-view structure containing clause nodes, Q&A nodes, penalty nodes, and their reference relationships; using a large model and a graph embedding model to jointly encode the graph to generate a potential vector representation of the regulatory clauses.

[0008] Sp2 Regulatory Response Field Construction: Based on the latent vectors of Sp1, a regulatory response field model is trained so that the state vector of the receiving agency and the reported feature vector output the compliance risk score under the corresponding regulatory interpretation.

[0009] Initial generation of Sp3 reporting strategy: Construct an agency state vector, and based on this state vector and the potential space of the regulatory response field, invoke a large model to generate an initial reporting strategy. This is represented as an executable reporting graph.

[0010] Sp4 Regulatory Adversary Agent Construction: Based on the regulatory reaction field and large model, a regulatory adversary agent is constructed, which generates a strict regulatory interpretation and searches for adversarial scenarios that are likely to lead to high-risk scores under the current reporting strategy and institutional status.

[0011] Sp5 Reporting Strategy – Optimization Against the Regulatory Opponent: Through multiple rounds of play between the reporting strategy and the regulatory opponent, unfavorable regulatory interpretations and input perturbations are applied in each round. The reporting strategy is then updated or transformed by the large model until a reporting strategy π robust to the most unfavorable regulatory interpretation is obtained. * .

[0012] SP6 Regulatory Elasticity Tensor Calculation: In Policy π * The study detours the main directions of potential regulatory space and assesses compliance risk responses, constructing a regulatory elasticity tensor to characterize the sensitivity of reporting results to changes in regulatory interpretation.

[0013] Sp7 Self-Evolving Reporting Update: When new or revised regulatory provisions are introduced, the new provisions are projected into the regulatory potential space, and the reporting graph is subjected to local structural transformation and parameter adjustment based on the regulatory flexibility modulus to generate a self-evolving reporting strategy π′.

[0014] Sp8 Reporting Output and Regulatory Simulation Interface Generation: Utilize strategy π′ to generate target period reporting procedures and reports, and export a regulatory simulation checker that can run independently on the regulatory side, used to verify the consistency of compliance risk assessment of the reporting results under different regulatory interpretation scenarios.

[0015] Preferably, the regulatory reaction field model described in Sp2 specifically includes:

[0016] Using an energy-based model or a diffusion model, the potential representation of regulatory provisions is jointly mapped with the institution's state vector and reporting feature vector to a regulatory energy value. The higher the energy value, the more likely the combination is to be regarded as high risk or violation by the regulator.

[0017] By applying a temperature parameter to the energy value and normalizing it, the probability distribution of compliance risk under different combinations of regulatory interpretations is obtained.

[0018] During the training process, the actual penalty results in penalty cases are used as a monitoring signal to update the parameters of the regulatory response field model, so that it outputs higher energy for high-risk reporting mode and lower energy for compliant reporting mode under various historical scenarios.

[0019] Preferably, the initial reporting strategy generated in Sp3 Represented as an executable delivery graph, the delivery graph includes:

[0020] Nodes represent operators such as data extraction, cleaning, aggregation, indicator calculation, and text generation;

[0021] Edges represent data flow and control flow dependencies between operators;

[0022] Each node is accompanied by a conditional vector of its input features in the latent space of the regulatory response field;

[0023] Furthermore, during the optimization process of Sp5, the reporting strategy is updated by replacing subgraphs, updating node parameters, and pruning paths in the reporting graph, resulting in an optimized reporting strategy π. * The graph structure remains executable and traceable.

[0024] Preferably, the adversary agent in Sp4 generates adversarial scenarios and counterexample reporting paths through the following steps:

[0025] A candidate "strict regulatory inquiry set" is generated using a large model for the current reporting task, and this inquiry set is mapped to a set of offset vectors in the regulatory potential space.

[0026] By combining the compliance risk gradient output by the regulatory reaction field model, a gradient ascent search is performed on the offset vector to obtain the regulatory interpretation combination that maximizes the energy value.

[0027] On the executable reporting graph, the input data is partially perturbed or replaced with alternative samples based on the above regulatory interpretation combination. Heuristic search or reinforcement learning is used to find the reporting path that maximizes the energy of compliance risk in the reporting path space, as adversarial scenarios and counterexamples.

[0028] Preferably, the reporting strategy in Sp5—the optimization process of monitoring the opponent's game—is a minimax game process, and its objective is:

[0029]

[0030] in, To report strategy parameters, A combination of regulatory interpretation and data manipulation. For a set of feasible strategies for regulating adversaries, For regulatory reaction field models, The loss function is based on energy value or expected compliance cost;

[0031] Furthermore, in actual optimization, an alternating update method is adopted: a fixed reporting strategy is used to optimize the parameters of the monitoring adversary agent and find the approximate worst-case scenario. Then fixed Update the reporting strategy parameters or adjust the reporting graph structure until the loss function converges within a preset number of rounds or the decrease is below a threshold.

[0032] Preferably, the process for constructing the regulatory resilience modulus in Sp6 includes:

[0033] Finite difference or automatic differentiation is performed on each principal component direction of the regulatory potential space, and a small perturbation is applied to each direction. While keeping the mechanism's state vector unchanged, the perturbed potential vector is compared with the convergence reporting strategy π. * By combining the input regulatory response field model, the change in compliance risk score is calculated. ;by and A local linear approximation of compliance risk to potential regulatory factors is constructed, forming a multidimensional tensor E, which is used to quantify the sensitivity of reporting output to changes in regulatory interpretation. The smaller the absolute value of the elements in E, the more robust the reporting strategy is in the corresponding direction.

[0034] Preferably, the self-evolutionary reporting updates in Sp7 include:

[0035] The newly added or revised regulatory provisions are encoded using a large model and projected onto the regulatory potential space to obtain a vector of provision changes. ;

[0036] Using the regulatory flexibility modulus E obtained from Sp6, estimate the direction and magnitude of the impact of clause changes on the current reporting strategy output, and calculate the recommended set of structural changes and parameter adjustment range;

[0037] Following the recommended changes, a low-rank structure transformation and local weight reparameterization are performed on the reporting graph to obtain the self-evolving reporting strategy. ;

[0038] Will Combined with the updated regulatory response field model, a rapid risk reassessment is performed on typical data scenarios. When the compliance risk scores of all scenarios are below the threshold, the self-evolution is confirmed to be successful; otherwise, a local game-playing fine-tuning is triggered instead of retraining from scratch.

[0039] Preferably, the regulatory simulation inspector exported by Sp8 includes:

[0040] A compressed version of the regulatory response field sub-model or a lightweight neural network derived from it is used to re-estimate the compliance risk score when the regulatory side inputs the agency status and reporting results;

[0041] A set of convergence reporting strategies π * Consistent indicator calculation and comparison rules are used to replay the calculation process of key indicators in a regulatory simulation environment;

[0042] An interface description file is used to specify how regulators inject potential bias vectors into the inspector when simulating different regulatory interpretation scenarios and read the corresponding compliance risk assessment results, thereby independently verifying the consistency between the financial institution's reporting strategy and the regulatory response field at the regulatory end.

[0043] Preferably, the hardware of the method includes:

[0044] The regulatory potential coding module is used to encode the regulatory interpretations, regulatory concerns, or situational inquiries input by regulatory agencies into potential vectors, and project these potential vectors onto potential space coordinates consistent with the regulatory response field model.

[0045] The lightweight regulatory reaction field replication module includes a lightweight neural network obtained by distilling the reaction field model. It is used to perform compliance risk energy assessment on the input reporting results without accessing the internal models and data of financial institutions, and outputs a simulated risk score of the reporting results under the current regulatory interpretation.

[0046] The replay reporting module is used to report according to the reporting strategy π. * Or its trimmed executable subgraph, recalculates the indicators based on the test data or publicly available sample data input by the regulatory agency, to simulate the main calculation path of the reporting strategy, and compares the recalculation results with the reporting results for consistency.

[0047] The adversarial scenario generation module is used to automatically generate a set of adversarial regulatory inspection scenarios that can be used by regulators based on the gradient information of the lightweight reaction field model and the potential spatial offset direction of regulation. These scenarios include: extreme interpretation combinations, structural change sensitive scenarios, and potential neighbors of historical penalty scenarios.

[0048] The regulatory verification interface is used to receive regulatory interpretation disturbances, penalty samples or specific inspection instructions input by regulators, and output simulated risk scores, recalculated indicators, consistency comparison results and corresponding adversarial scenarios, enabling regulatory agencies to independently verify the consistency and robustness between financial institutions' reporting strategies and the regulatory response field.

[0049] Beneficial effects

[0050] This invention provides an intelligent generation method for financial regulatory reports based on a large model. It has the following beneficial effects:

[0051] 1. This invention, through a combined framework of "regulatory reaction field + large model + game mechanism," enables reporting strategies to no longer rely on manual interpretation of regulations and manual compilation of report logic. Instead, it allows for automatic generation and optimization within the constraints of the potential space for regulatory interpretation, and achieves robustness to the most stringent regulatory interpretations through multiple rounds of maximal-minimal game theory. Compared to traditional reporting methods that rely on manual rules, this invention significantly reduces reporting errors caused by human misunderstanding biases. In real regulatory review scenarios, it makes the strategy more tolerant of abnormal inputs, changes in terminology, and regulatory inquiries, thereby improving the reporting compliance capabilities of institutions from an institutional perspective.

[0052] 2. This invention introduces a structural representation of the regulatory potential space and the regulatory elasticity tensor, allowing newly added or revised clauses to be directly mapped to change vectors in the potential space. The direction and magnitude of their impact on reporting results can be quickly inferred through the elasticity tensor. Based on this local sensitivity calculation mechanism, this invention enables reporting strategies to achieve self-evolutionary updates without retraining the entire game or reconstructing all reporting logic. Adaptation to new regulations requires only lightweight structural transformations and local parameter adjustments, significantly reducing system modification costs and compliance risks during regulatory change cycles.

[0053] 3. This invention provides a derivative, lightweight regulatory simulation checker, enabling regulatory agencies to independently simulate different interpretations, stress scenarios, or penalty-related neighboring scenarios within their own environments. It recalculates risk scores and key indicators using regulatory response field logic derived from that of financial institutions, achieving interpretive consistency between the reporting side and the regulatory side. This reproducible mechanism reduces regulatory communication costs, makes reporting logic transparent and verifiable, and allows regulators to efficiently identify potential non-compliance boundaries in reporting strategies, thereby promoting a higher-quality and more verifiable reporting ecosystem between regulators and financial institutions. Attached Figure Description

[0054] Figure 1 This is a flowchart of the intelligent generation method for financial regulatory reporting based on a large model, as described in this invention.

[0055] Figure 2 This is a system architecture diagram of the present invention;

[0056] Figure 3 This is a schematic diagram of the regulatory three-view structure and regulatory reaction field of the present invention;

[0057] Figure 4 This is a schematic diagram of the reporting strategy of the present invention – monitoring opponent's game and self-evolutionary update;

[0058] Figure 5 This is a screenshot of the system in operation according to the present invention;

[0059] Figure 6This is a screenshot of the operation of the regulatory reaction field system of the present invention. Detailed Implementation

[0060] The technical solutions of 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. Specific Implementation Example 1:

[0062] like Figures 1-6 As shown, the intelligent generation method for financial regulatory reports based on a large model includes the following steps:

[0063] Sp1 Regulatory Multi-Source Corpus Modeling: Obtain a regulatory corpus set including regulatory law texts, regulatory Q&A guidelines, and regulatory penalty cases. Preprocess the regulatory corpus and construct a three-view structure of regulation that includes clause nodes, Q&A nodes, penalty nodes, and the reference relationships between them. Use a large model and a graph embedding model to jointly encode the three-view structure of regulation to obtain the regulatory response vector representation of each regulatory clause.

[0064] Sp2 Regulatory Response Field Construction: Based on the regulatory response vector representation obtained in Sp1, a regulatory response field model is trained. The regulatory response field model receives the financial institution's state description vector and reporting feature vector, outputs the compliance risk score distribution of the reporting result under different regulatory interpretations, and forms a continuous potential representation space for regulatory provisions.

[0065] Initial generation of Sp3 reporting strategy: For the target reporting task, construct an institutional state vector from multi-source business data of financial institutions to describe the current asset and liability structure, risk exposure, and transaction behavior; call the large model to generate an initial reporting strategy containing reporting procedures, report structure, and text descriptions under the potential space conditions of the institutional state vector and the regulatory reaction field. The strategy is then encoded into an executable reporting graph.

[0066] Sp4 Regulatory Adversary Agent Construction: Based on the regulatory reaction field model and the large model, a regulatory adversary agent is constructed. The regulatory adversary agent receives the current reporting strategy and the institution's state vector, generates a set of the most stringent and reasonable interpretations of the regulatory provisions, and searches on the reporting graph for adversarial scenarios and counterexample reporting paths that can amplify the compliance risk score.

[0067] Sp5 Reporting Strategy – Optimization of Play Against the Supervisory Opponent: Multiple rounds of play are conducted between the reporting strategy obtained in Sp3 and the supervisory opponent agent in Sp4.

[0068] In each round, the regulatory adversary agent selects the most unfavorable regulatory interpretation and data disturbance scenario based on the compliance risk gradient output by the regulatory response field.

[0069] The reporting strategy generates a new reporting strategy through large model parameter updates or structural transformations. This minimizes the compliance risk score under the most unfavorable scenario described above.

[0070] Repeat the iteration until a reporting strategy π that is robust to the worst-case regulatory interpretation is obtained under the preset convergence condition. * .

[0071] SP6 Regulatory Elasticity Tensor Calculation:

[0072] After obtaining the convergence reporting strategy π * Subsequently, small perturbations are made to each major direction in the regulatory potential space. The perturbed potential vectors are then input into the regulatory response field model and the reporting strategy combination. The sensitivity of the reporting results to each regulatory potential factor is calculated to form the regulatory elasticity modulus. The regulatory elasticity tensor characterizes the directional change of the reporting strategy output when the potential regulatory interpretation changes.

[0073] SP7 Self-Evolution Report Update:

[0074] In the actual operation phase, when new or revised regulatory provisions are detected, the new provisions are encoded and projected into the regulatory potential space. Based on the regulatory elasticity modulus of Sp6, the minimum adjustment direction and magnitude of the change in the reporting strategy output space are estimated. Without re-training the entire game, the reporting graph is structurally compressed and locally reparameterized to obtain the self-evolved reporting strategy π′. The regulatory reaction field is used to verify that π′ still meets the preset compliance risk threshold under the new provisions.

[0075] Sp8 Report Output and Regulatory Simulation Interface Generation:

[0076] Based on the self-evolving reporting strategy π′, the financial data for the target period is used to generate reporting procedures and reports, outputting reporting results including report data, textual descriptions, and key compliance indicators. At the same time, a corresponding regulatory simulation checker is exported. This checker is a lightweight model or set of rules that can run independently on the regulatory side. It is used to reproduce the regulatory response field for compliance risk assessment of the reporting results when the regulatory agency simulates different interpretation scenarios, thereby achieving consistent verification between the reporting results and the regulatory simulation.

[0077] The regulatory reaction field model in Sp2 specifically includes:

[0078] Using an energy-based model or a diffusion model, the potential representation of regulatory provisions is jointly mapped with the institution's state vector and reporting feature vector to a regulatory energy value. The higher the energy value, the more likely the combination is to be regarded as high risk or violation by the regulator.

[0079] By applying a temperature parameter to the energy value and normalizing it, the probability distribution of compliance risk under different combinations of regulatory interpretations is obtained.

[0080] During the training process, the actual penalty results in penalty cases are used as a monitoring signal to update the parameters of the regulatory response field model, so that it outputs higher energy for high-risk reporting mode and lower energy for compliant reporting mode under various historical scenarios.

[0081] Initial reporting strategy generated in SP3 Represented as an executable submission diagram, the submission diagram includes:

[0082] Nodes represent operators such as data extraction, cleaning, aggregation, indicator calculation, and text generation;

[0083] Edges represent data flow and control flow dependencies between operators;

[0084] Each node is accompanied by a conditional vector of its input features in the latent space of the regulatory response field;

[0085] Furthermore, during the optimization process of Sp5, the reporting strategy is updated by replacing subgraphs, updating node parameters, and pruning paths in the reporting graph, resulting in an optimized reporting strategy π. * The graph structure remains executable and traceable.

[0086] In SP4, the regulatory adversary agent generates adversarial scenarios and counterexample reporting paths through the following steps:

[0087] A candidate "strict regulatory inquiry set" is generated using a large model for the current reporting task, and this inquiry set is mapped to a set of offset vectors in the regulatory potential space.

[0088] By combining the compliance risk gradient output by the regulatory reaction field model, a gradient ascent search is performed on the offset vector to obtain the regulatory interpretation combination that maximizes the energy value.

[0089] On the executable reporting graph, the input data is partially perturbed or replaced with alternative samples based on the above regulatory interpretation combination. Heuristic search or reinforcement learning is used to find the reporting path that maximizes the energy of compliance risk in the reporting path space, as adversarial scenarios and counterexamples.

[0090] The reporting strategy in SP5—the optimization process of monitoring the opponent's game—is a minimax game process with the following objective:

[0091]

[0092] in, To report strategy parameters, A combination of regulatory interpretation and data manipulation. For a set of feasible strategies for regulating adversaries, For regulatory reaction field models, The loss function is based on energy value or expected compliance cost;

[0093] Furthermore, in actual optimization, an alternating update method is adopted: a fixed reporting strategy is used to optimize the parameters of the monitoring adversary agent and find the approximate worst-case scenario. Then fixed Update the reporting strategy parameters or adjust the reporting graph structure until the loss function converges within a preset number of rounds or the decrease is below a threshold.

[0094] The process of constructing the regulatory resilience modulus in SP6 includes:

[0095] Finite difference or automatic differentiation is performed on each principal component direction of the regulatory potential space, and a small perturbation is applied to each direction. While keeping the mechanism's state vector unchanged, the perturbed potential vector is compared with the reporting strategy. By combining the input regulatory response field model, the change in compliance risk score is calculated. ;by and Construct a local linear approximation of compliance risk to potential regulatory factors, forming a multidimensional tensor. This is used to quantify the sensitivity of reporting outputs to changes in regulatory interpretation. The smaller the absolute value of the element, the more robust the reporting strategy is in the corresponding direction.

[0096] The self-evolution reporting updates in SP7 include:

[0097] The newly added or revised regulatory provisions are encoded using a large model and projected onto the regulatory potential space to obtain a vector of provision changes. ;

[0098] Regulatory elasticity modulus obtained using Sp6 Estimate the direction and magnitude of the impact of the changes in the terms on the current reporting strategy output, and calculate the recommended set of structural changes and the range of parameter adjustments;

[0099] Following the recommended changes, a low-rank structure transformation and local weight reparameterization are performed on the reporting graph to obtain the self-evolving reporting strategy. ;

[0100] Will Combined with the updated regulatory response field model, a rapid risk reassessment is performed on typical data scenarios. When the compliance risk scores of all scenarios are below the threshold, the self-evolution is confirmed to be successful; otherwise, a local game-playing fine-tuning is triggered instead of retraining from scratch.

[0101] The regulatory simulation inspector exported by SP8 includes:

[0102] A compressed version of the regulatory response field sub-model or a lightweight neural network derived from it is used to re-estimate the compliance risk score when the regulatory side inputs the agency status and reporting results;

[0103] A set of reporting strategies Consistent indicator calculation and comparison rules are used to replay the calculation process of key indicators in a regulatory simulation environment;

[0104] An interface description file is used to specify how regulators inject potential bias vectors into the inspector when simulating different regulatory interpretation scenarios and read the corresponding compliance risk assessment results, thereby independently verifying the consistency between the financial institution's reporting strategy and the regulatory response field at the regulatory end.

[0105] The hardware for the method includes:

[0106] The regulatory potential coding module is used to encode the regulatory interpretations, regulatory concerns, or situational inquiries input by regulatory agencies into potential vectors, and project these potential vectors onto potential space coordinates consistent with the regulatory response field model.

[0107] The lightweight regulatory reaction field replication module includes a lightweight neural network obtained by distilling the reaction field model. It is used to perform compliance risk energy assessment on the input reporting results without accessing the internal models and data of financial institutions, and outputs a simulated risk score of the reporting results under the current regulatory interpretation.

[0108] The replay reporting module is used to report according to the reporting strategy. Or its trimmed executable subgraph, recalculates the indicators based on the test data or publicly available sample data input by the regulatory agency, to simulate the main calculation path of the reporting strategy, and compares the recalculation results with the reporting results for consistency.

[0109] The adversarial scenario generation module is used to automatically generate a set of adversarial regulatory inspection scenarios that can be used by regulators based on the gradient information of the lightweight reaction field model and the potential spatial offset direction of regulation. These scenarios include: extreme interpretation combinations, structural change sensitive scenarios, and potential neighbors of historical penalty scenarios.

[0110] The regulatory verification interface is used to receive regulatory interpretation disturbances, penalty samples or specific inspection instructions input by regulators, and output simulated risk scores, recalculated indicators, consistency comparison results and corresponding adversarial scenarios, enabling regulatory agencies to independently verify the consistency and robustness between financial institutions' reporting strategies and the regulatory response field. Specific Implementation Example 2:

[0112] Based on the technical solution of Specific Embodiment 1, the following applications are further provided:

[0113] A banking group is required to submit regulatory indicators such as LCR (Liquidity Coverage Ratio) and NSFR (Net Stable Funding Ratio) to regulatory agencies monthly. The current practice is that the IT department has a fixed reporting procedure, and when regulatory requirements change, the code needs to be modified manually and regression tested manually, which is time-consuming and prone to errors.

[0114] The system is deployed in the bank's internal data center and includes the following logical layers:

[0115] Data Layer: ODS, DW, Risk Engine Results, Transaction Logs, etc. ODS stands for Operational Data Store, which is the lowest level of the bank's internal data system and the closest to the business system, providing real-time or near-real-time data. DW stands for Data Warehouse, which is a data storage system built on top of ODS, cleaned, integrated, and thematically modeled, and is geared towards the middle platform and reporting.

[0116] Model service layer: Large Language Model (LLM) service, Graph Neural Network service, Regulatory Reaction Field Energy Model service.

[0117] Reporting game engine: reporting strategy generator, monitoring opponent agent, minimax trainer.

[0118] Self-evolution engine: Regulatory elastic tensor calculator, incremental update executor.

[0119] Regulatory simulation devices (deployed on the regulatory side or in a sandbox environment).

[0120] Sp1: Modeling of Regulatory Multi-Source Corpus:

[0121] 1. Data preparation: Regulatory texts such as the "Guidelines for Liquidity Risk Management" and the "Capital Management Measures", divided into sections according to clauses.

[0122] Regulatory Q&A and Explanatory Documents: Q&A documents and explanations issued by regulators.

[0123] Penalty Case: Penalty Notice Text, including a description of the violation, the clauses violated, and the penalty result.

[0124] These texts are stored in the `reg_corpus` table. Example of fields:

[0125] id:UUID

[0126] type:{ARTICLE,Q&A,CASE}

[0127] ref_law_id: Law number or clause number

[0128] Original text

[0129] meta: JSON (including time, scope of application, amount of fine, etc.)

[0130] 2. Construct a "three-view structure for supervision":

[0131] Construct a graph G=(V,E) with three types of nodes:

[0132] Clause Node V article One node for each regulatory clause;

[0133] Question and Answer Node V qa Each regulatory Q&A has one node;

[0134] Penalty Node V case Each penalty notice has one node.

[0135] Edge types include:

[0136] CITATION: Citation clauses in penalty case examples;

[0137] INTERPRET: Questions and answers that provide an interpretation of a clause;

[0138] SIMILAR: Semantic similarity between terms (established through text similarity thresholds).

[0139] In actual construction:

[0140] Encode each text using a text vector model (e.g., a 768-dimensional sentence vector);

[0141] Calculate the cosine similarity between clauses, between questions and answers and clauses, and between cases and clauses;

[0142] Create SIMILAR edges if the threshold is exceeded (e.g., 0.75); create CITATION / INTERPRET edges based on keywords such as "in accordance with a certain clause" or "refer to Article X" in the meta tag.

[0143] Nodes are stored in the table reg_-Nodes, edges are stored in the table reg_edges, and synchronized to a graph database (such as Neo4j).

[0144] 3. Jointly encoded as a "regulatory response vector":

[0145] We use a graph neural network (GNN) + large model embedding framework:

[0146] Initial text embedding:

[0147] Using a sentence vector model specifically tuned for legal / financial applications, the text of each node is embedded as... (e.g., d_text=1024).

[0148] Structure encoding:

[0149] Using graph neural networks (such as GraphSAGE or GAT), perform several rounds of message propagation on graph G:

[0150] In each round of propagation, the vectors of adjacent nodes are aggregated, and the weight of the edge type is taken into account (the weight of referenced edges is greater than that of ordinary similar edges).

[0151] Obtain the structure enhancement vector for each node. (e.g., d_gnn=256).

[0152] splicing and dimensionality reduction:

[0153] Node final vector W is a linear projection matrix, and the output is a 512-dimensional vector, denoted as the regulatory response vector. .

[0154] Result: Each regulatory provision has one The Q&A and case nodes also have corresponding vectors, which form the basis for the subsequent "regulatory response field".

[0155] Sp2: Regulatory Response Field Model

[0156] 1. Model Form: The energy-based regulatory reaction field defines an energy function:

[0157] in: : Institutional state vector (the overall situation of the bank in the current period); : Feature vector of the reported results; : Potential vectors for regulatory interpretation (derived from a combination of terms / inquiries); Model parameters.

[0158] Intuitive meaning: The larger the value, the more significant the regulatory interpretation. state Reporting results "In this combination, regulators are more likely to consider it high-risk or illegal."

[0159] Input feature design:

[0160] State vector This includes asset size, loan and deposit structure, foreign exchange position, historical LCR / NSFR volatility, etc., all of which are assembled into a 256-dimensional vector.

[0161] Reporting result vector This includes the current period's LCR / NSFR values, various sub-indicators (balance of high-quality liquid assets, net cash outflows, etc.), and structural characteristics such as asset concentration and maturity mismatch, totaling 128 dimensions.

[0162] Regulatory interpretation : Select a set of clauses / inquiries from the regulatory three-view diagram, and determine its regulatory response vector. By performing a weighted average or attention fusion, we can obtain... .

[0163] 3. The model structure uses a multi-layer MLP or Transformer:

[0164] input:

[0165] layer1: 1024 ReLU

[0166] layer2: 512 ReLU

[0167] layer3: 128 ReLU

[0168] output:

[0169] To enhance expressive power, residual connections or normalization can be added to the intermediate layers.

[0170] 4. Training labels and loss function:

[0171] Constructing training samples:

[0172] Positive samples (compliant):

[0173] Reports that have not been penalized in historical submission records and have no issues during on-site regulatory inspections;

[0174] Marked as "low energy target", hoping for low E.

[0175] Negative samples (high risk / violation):

[0176] The reporting records (or simulated reconstructions) corresponding to the penalty cases, and the related clauses z;

[0177] Submitted samples with LCR / NSFR significantly lower than regulatory requirements or with abnormal structures;

[0178] Marked as a "high-energy target", hoping high.

[0179] Using contrastive loss:

[0180] in Provide the energy interval between positive and negative samples; additionally, a contrastive learning loss can be added to widen the energy difference between different y values ​​under the same state.

[0181] After training:

[0182] Given (s,y,z), we can obtain E;

[0183] Then through Mapped to risk probability;

[0184] This is the "regulatory response field," which will be used for game theory and resilience analysis later on.

[0185] SP3: Initial generation of reporting strategy:

[0186] 1. Construction of the mechanism state vector s:

[0187] Extracting from the data warehouse:

[0188] Asset side: Distributed by product, term, and currency;

[0189] Liabilities side: Same as above;

[0190] Derivatives, off-balance-sheet items, etc.;

[0191] Historical regulatory indicator series (LCR / NSFR, etc. over the past 12 months);

[0192] Market data (interest rates, exchange rates).

[0193] These features are assembled into a fixed-dimensional vector s (e.g., 256-dimensional), which is achieved through standardization and feature selection.

[0194] 2. Definition of submitted figure:

[0195] We define the basic elements of the executable reporting graph RGraph:

[0196] Node type:

[0197] Extract-Node: Extracts data from a specified table / field;

[0198] Filter-Node: Filter by conditions (e.g., remove related parties, specific currencies);

[0199] Aggregate-Node: Aggregates by a certain dimension (sum, avg, max, etc.);

[0200] Compute-Node: Formula for calculating metrics;

[0201] Explain-Node: Generates text descriptions;

[0202] Edges: indicate data dependencies and execution order. RGraph guarantees that it is a DAG.

[0203] Each node is represented in JSON format.

[0204] {

[0205] "type":"Aggregate-Node",

[0206] "input":"exposures_after_filter",

[0207] "group_by":["currency"],

[0208] "metric":"sum",

[0209] "output":"liq_assets_by_ccy"

[0210] }

[0211] 3. LLM generates initial reporting strategy :

[0212] Use a large language model (e.g., 13B / 30B parameters) + a specific prompt:

[0213] enter:

[0214] Summary of regulatory provisions;

[0215] Key summary of the institution's status (e.g., "high foreign currency position ratio, large short-term interbank liabilities ratio");

[0216] Historical LCR / NSFR calculation rule template.

[0217] Output:

[0218] A JSON representation of an RGraph;

[0219] Natural language descriptions of key steps (for review).

[0220] The correctness of JSON syntax is ensured by "syntactic constraint decoding", and the validity of node types is ensured by schema validation.

[0221] After generation, run dryrun once in the system and check:

[0222] Are all the required data sources complete?

[0223] Check whether the algorithm logic is executable and whether there are any circular dependencies or data type errors.

[0224] This is the initial reporting strategy. .

[0225] Sp4: Regulatory adversary agents:

[0226] The goal of monitoring adversary agents: in the "current reporting strategy" Under the state "s", find a set of regulatory interpretations and data disturbance Let energy Make it as large as possible.

[0227] 1. Generate a "Strict Regulatory Inquiry Set": Use LLM to generate a set of possible inquiries / explanations for LCR / NSFR on the regulatory corpus, such as:

[0228] "Including certain types of interbank deposits in adverse scenario cash outflows";

[0229] "Requirements to increase the risk weight or reduce the proportion of certain assets," etc.

[0230] Each query text is mapped into a vector by the aforementioned encoder, then fed into the GNN to obtain the latent vector offset. .

[0231] 2. Regulatory interpretation combination search (within the potential space):

[0232] The basis of a given set of terms Regulatory adversary intelligent agents seek out opportunities through gradient ascent. :

[0233] initialization = ;

[0234] Calculate energy ;

[0235] calculate The gradient is obtained through automatic differentiation;

[0236] renew Or in the form of Adam;

[0237] Constrain z to prevent it from deviating too far from the original clause distribution (e.g., L2 norm restriction) to avoid unreasonable interpretations;

[0238] After K iterations, the regulatory interpretation z is obtained. * .

[0239] 3. Reporting path counterexample search:

[0240] To find "more regulatory sensitive" paths on RGraph, one could do the following:

[0241] Enumerate combinations of key nodes (such as specific asset classification methods or critical caliber classifications) to construct a finite set;

[0242] Define a policy space: allow switching between a few optional logics on these nodes, for example:

[0243] Whether a certain type of deposit is included in stable funds or not;

[0244] Include certain types of interbank assets in high-quality liquid assets vs. partially included.

[0245] Use reinforcement learning (RL) or heuristic search:

[0246] Status: Current RGraph switch configurations;

[0247] Action: Switch one of the switches;

[0248] Reward: Under the current state, the energy of the regulatory reaction field. The magnitude of the increase;

[0249] Objective: Find the configuration that maximizes energy as a "counterexample path".

[0250] Output: A set of "extreme regulatory interpretations" +Reporting Counterexamples Configuration "" as a regulatory counterpart strategy.

[0251] Sp5: Reporting Strategy – Monitoring Opponent's Extreme Minimum Play:

[0252] 1. Formal approximation of the game objective:

[0253] π: Reporting strategy (corresponding to RGraph and node parameters);

[0254] :A combination of regulatory interpretations and data / logic disturbances;

[0255] :Depend on The resulting regulatory potential vector.

[0256] Algorithm flow (single state s)

[0257] initialization .

[0258] Fort=1…T (training rounds):

[0259] fixed Optimize regulatory counterparts :

[0260] Run the potential space + path search of SP4 to find the approximate worst case. .

[0261] fixed Update reporting strategy :

[0262] Calculate the current reporting results ;

[0263] Calculate energy loss L= ;

[0264] For the "node parameterization part" (such as classification threshold, weighting coefficients, etc.), gradient descent is used directly;

[0265] For the reported graph structure section, use structure search:

[0266] Record the scores of multiple candidate RGraphs and employ a strategy similar to Neural Architecture Search (NAS) to retain results with lower energy.

[0267] For high-energy structures, try "replacing a subgraph with a new subgraph generated by LLM", then calculate the energy. If the energy decreases, accept the result.

[0268] Convergence condition: The decrease in L is less than ε for several consecutive rounds, or the energy is lower than a certain target threshold.

[0269] Ultimately, we obtain a reporting strategy π that maintains a low E for the worst-case scenario δ under the current state s. * .

[0270] 3. Batch training:

[0271] For multiple historical states Batch execution of the above game is similar to dialogue strategy training. Each update can share a portion of the reported graph structure template, and multi-task learning is used for node parameters, thereby obtaining a set of π that is robust to multiple states. * .

[0272] Sp6: Calculation of the regulatory elasticity tensor:

[0273] 1. Selection of potential spatial directions: Perform PCA or principal component analysis on the potential space z (512 dimensions) of regulatory provisions to obtain the top M (e.g., 20) principal directions. .

[0274] The elasticity calculation process for each direction :

[0275] calculate ( For very small step sizes, such as 0.1).

[0276] Fixed state Reporting strategy Calculate the corresponding energy , .

[0277] Changes in metric outputs (such as LCR / NSFR) can also be calculated simultaneously. Definition:

[0278] For more detail, it can be extended to two dimensions to form a tensor. .

[0279] The result is a vector / tensor that tells us how sensitive the risk energy is to changes along a certain direction of regulatory interpretation e_i.

[0280] Storage and use will (or some simplification thereof) and the corresponding direction Store it in the "Regulatory Elasticity Library" for use by SP7.

[0281] SP7: Self-evolutionary reporting update assumes that the regulatory authorities have released a new document that adjusts the LCR definition.

[0282] The new clause text is encoded into a vector using the same text encoder and GNN. ;

[0283] Obtain new potential vectors ;

[0284] Calculate the clause change vector (Weighted by multiple terms).

[0285] Using the previous SP6 Base and ,Will Expand on these bases:

[0286] ;

[0287] Estimate the change in risk energy: If If it is very small, then the existing strategy π is considered to be... * It remains relatively safe under the new regulations, requiring only minor adjustments; if If the size is large, then a larger-scale structural adjustment is needed.

[0288] The report graph's structural transformation and local reparameterization engine are based on and Distribution:

[0289] Find the right The areas that contributed the most, corresponding to the sensitive operators in RGraph:

[0290] Label the aggregation nodes related to a certain asset classification "Sensitive nodes".

[0291] Perform low-rank structural transformations near sensitive nodes: add / replace subgraphs: generate new local logic using LLM under new clause constraints;

[0292] Adjust node parameters: Adjust weights, thresholds, and classification boundaries.

[0293] Other nodes remain unchanged, thus avoiding a full graph reconstruction.

[0294] Use the updated reaction field model (trained with new terms or fine-tuned) on a typical state sample set { Verification will be performed:

[0295] For each ,calculate ;

[0296] If the energy of all samples is below the threshold, then the self-evolution strategy π′ is accepted.

[0297] Otherwise, only a small-scale game is started to fine-tune the local area corresponding to the failed sample (instead of starting from scratch).

[0298] In this way, after the new regulations are released, the system can automatically complete the policy migration within a short period of time.

[0299] SP8: Output and Regulatory Simulation Inspector

[0300] 1. Reporting results are generated during the official reporting cycle:

[0301] Get the current state , and the latest strategy π′;

[0302] Execute RGraph to obtain regulatory metrics such as LCR / NSFR, itemized tables, and automatically generated text descriptions (Explain-Node output).

[0303] Simultaneously, the regulatory reaction field model is invoked, under the default regulatory interpretation. The system calculates the energy and risk probability and outputs the "self-check result".

[0304] The report package includes:

[0305] Structured reports (XML, Excel);

[0306] Reporting procedure instructions (human-readable);

[0307] Self-assessment compliance risk score and key explanatory factors.

[0308] 2. Regulatory Simulation Inspector: Generates a lightweight sub-model of the regulatory reaction field on the agency side, and uses knowledge distillation to... Extract:

[0309] Reduce the number of layers and parameters (e.g., from a 3-layer MLP to a 2-layer MLP) to ensure fast operation even on the regulatory side;

[0310] Reduce input features: Only retain the y and s features that have the greatest impact on risk assessment.

[0311] Simultaneously generated:

[0312] Indicator replay rules: Simple calculation logic allows regulators to recalculate LCR / NSFR using their own data or samples;

[0313] Interface Description: Input is specified in JSON format, including status summary, reporting indicators, and regulatory interpretation offset vector. wait;

[0314] Outputs: Risk score, energy value, and contribution of key features.

[0315] After the simulation device is imported into the regulatory system, the following can be achieved:

[0316] Input different scenarios (such as stress situations or extreme interpretations) and see how the risk score changes;

[0317] Whether the reporting strategy of the verification agency is sound from a regulatory perspective.

[0318] Training update cycle:

[0319] Regulatory response field model: initial training + quarterly fine-tuning using the latest penalty cases.

[0320] Reporting strategy game training: The initial time is relatively long (several hours or days), but subsequent incremental adjustments are based on elastic tensors, and the cycle can be shortened to days or hours.

[0321] Performance monitoring metrics:

[0322] Historical data is reviewed, and the deviation between the reported result and the previously reported result is less than a certain threshold.

[0323] Recall rate of regulatory response field on historical penalty samples (proportion of high-risk samples with high energy ranking).

[0324] After the new regulations were released, the adaptation time and manpower required to switch from the old strategy to π′;

[0325] Pass rate of regulatory sandbox testing (verified using simulation devices).

[0326] Security and Auditing:

[0327] All RGraph versions generated by LLM have a version number and a timestamp;

[0328] Audit records should be kept for each structural adjustment and parameter update;

[0329] In the event of a serious deviation, you can roll back to the previous policy version.

[0330] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0331] 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 method for intelligently generating financial regulatory reports based on a large model, characterized in that: The process includes the following steps: Sp1 Regulatory Multi-Source Corpus Modeling: Obtain regulatory corpus texts, regulatory Q&A guidelines, and regulatory penalty cases; preprocess them and construct a three-view structure of regulation containing clause nodes, Q&A nodes, penalty nodes, and their referencing relationships; use a large model and a graph embedding model to jointly encode the graph and generate a latent vector representation of the regulatory clauses; Sp2 Regulatory Response Field Construction: Based on the potential vectors of Sp1, a regulatory response field model is trained. The regulatory response field model adopts an energy-based model or a diffusion model, which jointly maps the potential representation of regulatory provisions with the institution's state vector and reporting feature vector into a regulatory energy value. This value enables the institution to receive the institution's state vector and reporting feature vector and output the corresponding compliance risk score under the regulatory interpretation. The institution's state vector includes asset size, deposit and loan structure, foreign exchange position, and historical LCR / NSFR volatility. The reporting feature vector includes the current period's LCR / NSFR value, sub-indicators, and structural features. Sp3 Reporting Strategy Initial Generation: Construct an agency state vector, and based on the state vector and the potential space of the regulatory response field, call the large model to generate the initial reporting strategy π0, which is represented as an executable reporting graph; Sp4 Regulatory Adversary Agent Construction: Based on the regulatory reaction field and large model, a regulatory adversary agent is constructed to generate a strict regulatory interpretation and search for adversarial scenarios that are likely to lead to high-risk scores under the current reporting strategy and institutional status. Sp5 Reporting Strategy – Optimization Against the Regulatory Opponent: Through multiple rounds of play between the reporting strategy and the regulatory opponent, unfavorable regulatory interpretations and input perturbations are applied in each round. The reporting strategy is then updated or transformed by the large model until a reporting strategy π robust to the most unfavorable regulatory interpretation is obtained. * ; SP6 Regulatory Elasticity Tensor Calculation: In Policy π * The following approach is used to perturb the potential direction of regulation and assess compliance risk response, constructing a regulatory elasticity tensor to characterize the sensitivity of reporting results to changes in regulatory interpretation. Sp7 Self-Evolving Reporting Update: When new or revised regulatory provisions are added, the new provisions are projected onto the regulatory potential space, and the reporting graph is subjected to local structural transformation and parameter adjustment based on the regulatory flexibility modulus to generate a self-evolving reporting strategy π′. Sp8 Reporting Output and Regulatory Simulation Interface Generation: Utilize strategy π′ to generate target period reporting procedures and reports, and export a regulatory simulation checker that can run independently on the regulatory side, used to verify the consistency of compliance risk assessment of the reporting results under different regulatory interpretation scenarios.

2. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, The regulatory reaction field model in Sp2 specifically includes: The higher the energy value of the regulatory energy value, the more likely the regulatory reaction field model is to be regarded as high-risk or non-compliant by the regulator. By applying a temperature parameter to the energy value and normalizing it, the probability distribution of compliance risk under different regulatory interpretations and regulatory response field models is obtained. During the training process, the actual penalty results in penalty cases are used as a monitoring signal to update the parameters of the regulatory response field model, so that it outputs higher energy for high-risk reporting mode and lower energy for compliant reporting mode under various historical scenarios.

3. The intelligent generation method for financial regulatory reporting based on a large model according to claim 2, characterized in that, The initial reporting strategy π0 generated in Sp3 is represented as an executable reporting graph, which includes: Nodes represent data extraction, cleaning, aggregation, indicator calculation, and text generation operators; Edges represent data flow and control flow dependencies between operators; Each node is accompanied by a conditional vector of its input features in the latent space of the regulatory response field; Furthermore, during the optimization process of Sp5, the reporting strategy is updated by replacing subgraphs, updating node parameters, and pruning paths in the reporting graph, resulting in an optimized reporting strategy π. * The graph structure remains executable and traceable.

4. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, In SP4, the regulatory adversary agent generates adversarial scenarios and counterexample reporting paths through the following steps: A candidate "strict regulatory inquiry set" is generated using a large model for the current reporting task, and this inquiry set is mapped to a set of offset vectors in the regulatory potential space. By combining the compliance risk gradient output by the regulatory reaction field model, a gradient ascent search is performed on the offset vector to obtain the regulatory interpretation combination that maximizes the energy value. On the executable reporting graph, the input data is partially perturbed or replaced with alternative samples based on the above regulatory interpretation combination. Heuristic search or reinforcement learning is used to find the reporting path that maximizes the energy of compliance risk in the reporting path space, as adversarial scenarios and counterexamples.

5. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, The reporting strategy in SP5—the optimization process of monitoring the opponent's game—is a minimax game process with the following objective: ； in, To report strategy parameters, A combination of regulatory interpretation and data manipulation. For a set of feasible strategies for regulating adversaries, For regulatory reaction field models, The loss function is based on energy value or expected compliance cost; This is the institution's state vector, representing the core risks and business status of a financial institution during a specific reporting period; Indicates the reporting strategy parameters The reporting output results are generated under the guidance of regulatory interpretation and the combined influence of data disturbance δ. Furthermore, in actual optimization, an alternating update method is adopted: a fixed reporting strategy is used to optimize the parameters of the monitoring adversary agent and find the approximate worst-case scenario. Then fixed Update the reporting strategy parameters or adjust the reporting graph structure until the loss function converges within a preset number of rounds or the decrease is below a threshold.

6. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, The process of constructing the regulatory resilience modulus in SP6 includes: Finite difference or automatic differentiation is performed on each principal component direction of the regulatory potential space, and a small perturbation is applied to each direction. While keeping the mechanism's state vector unchanged, the perturbed potential vector is compared with the convergence reporting strategy π. * By combining the input regulatory response field model, the change in compliance risk score is calculated. ;by and A local linear approximation of compliance risk to potential regulatory factors is constructed, forming a multidimensional tensor E, which is used to quantify the sensitivity of reporting output to changes in regulatory interpretation. The smaller the absolute value of the elements in E, the more robust the reporting strategy is in the corresponding direction.

7. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, The self-evolution reporting updates in SP7 include: The newly added or revised regulatory provisions are encoded using a large model and projected onto the regulatory potential space to obtain a vector of provision changes. ; Using the regulatory flexibility modulus E obtained from Sp6, estimate the direction and magnitude of the impact of clause changes on the current reporting strategy output, and calculate the recommended set of structural changes and parameter adjustment range; Following the recommended changes, a low-rank structure transformation and local weight reparameterization are performed on the reporting graph to obtain the self-evolving reporting strategy. ; Will Combined with the updated regulatory response field model, a rapid risk reassessment is performed on typical data scenarios. When the compliance risk scores of all scenarios are below the threshold, the self-evolution is confirmed to be successful; otherwise, a local game-playing fine-tuning is triggered instead of retraining from scratch.

8. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, The regulatory simulation inspector exported by SP8 includes: A compressed version of the regulatory response field sub-model or a lightweight neural network derived from it is used to re-estimate the compliance risk score when the regulatory side inputs the agency status and reporting results; A set of convergence reporting strategies π * Consistent indicator calculation and comparison rules are used to replay the calculation process of key indicators in a regulatory simulation environment; An interface description file is used to specify how regulators inject potential bias vectors into the inspector when simulating different regulatory interpretation scenarios and read the corresponding compliance risk assessment results, thereby independently verifying the consistency between the financial institution's reporting strategy and the regulatory response field at the regulatory end.

9. The intelligent generation method for financial regulatory reporting based on a large model according to claim 1, characterized in that, The hardware for the method includes: The regulatory potential coding module is used to encode the regulatory interpretations, regulatory concerns, or situational inquiries input by the regulatory agency into potential vectors, and project the potential vectors onto potential space coordinates consistent with the regulatory response field model in the method. The lightweight regulatory reaction field replication module includes a lightweight neural network obtained by distilling the reaction field model. It is used to perform compliance risk energy assessment on the input reporting results without accessing the internal models and data of financial institutions, and outputs a simulated risk score of the reporting results under the current regulatory interpretation. The replay reporting module is used to report according to the reporting strategy π. * Or its trimmed executable subgraph, recalculates the indicators based on the test data or publicly available sample data input by the regulatory agency, to simulate the calculation path of the reporting strategy, and compares the recalculation results with the reporting results for consistency. The adversarial scenario generation module is used to automatically generate a set of adversarial regulatory inspection scenarios that can be used by regulators based on the gradient information of the lightweight reaction field model and the potential spatial offset direction of regulation. These scenarios include: extreme interpretation combinations, structural change sensitive scenarios, and potential neighbors of historical penalty scenarios. The regulatory verification interface is used to receive regulatory interpretation disturbances, penalty samples or inspection instructions input by regulators, and output simulated risk scores, recalculated indicators, consistency comparison results and corresponding adversarial scenarios, enabling regulatory agencies to independently verify the consistency and robustness between financial institutions' reporting strategies and the regulatory response field.

Images