A multi-role large language model simulation conclusion quantitative fusion method and system based on synthetic anchors

Abstract

The application discloses a multi-role large language model simulation conclusion quantitative fusion method and system based on a synthetic anchor point, and belongs to the technical field of artificial intelligence simulation. A synthetic driving index with constant value is introduced in a base period index set, qualitative index influence judgment output by a large language model role is converted into a conduction rule starting from the index, and the synthetic index is removed after quantitative simulation is completed by inputting a deterministic numerical value simulation engine; a role confidence label is converted into a standard deviation coefficient according to a preset mapping table, and the most conservative value is taken when multiple roles jointly judge; the application does not depend on additional large language model calling, votes are conducted by counting whether the change direction given by each role to multiple indexes is consistent, and a correlation coefficient between indexes is inferred. The application solves the technical problems of qualitative simulation conclusion and quantitative engine disconnection, uncertainty measurement and judgment confidence disconnection, and high correlation matrix estimation cost, and can be used for an enterprise operation decision support platform.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and computer data simulation technology. Specifically, it relates to a method and system for converting qualitative judgments generated by multi-role large language model simulation into quantifiable uncertainty intervals. It is particularly suitable for intelligent decision-making platforms in enterprise business decision-making scenarios that need to integrate the multi-agent game deduction conclusions of large language models with deterministic numerical simulation engines. Background Technology

[0002] In enterprise decision support scenarios, users often need to assess the impact of external events (such as competitor price reductions, cost shocks, and major customer churn) on operational metrics. Traditional numerical simulation methods (such as Monte Carlo propagation simulation based on historical ratios) excel at quantifying confidence intervals for known numerical assumptions, but they cannot handle qualitative judgments requiring multi-agent game reasoning, such as "how stakeholders will react after an event occurs." On the other hand, multi-role simulation methods based on large language models are good at generating qualitative judgments of multi-agent game theory, but their output is free text or loosely structured natural language conclusions, lacking rigorous uncertainty quantification and correlation modeling capabilities, and cannot be directly used in scenarios with high quantitative rigor, such as financial risk assessment. The main solutions in the existing technology include: Option 1: Pure numerical simulation. Users manually convert the event into several numerical hypotheses (e.g., "customer numbers decreased by 30%)" and input them into a deterministic numerical simulation engine to obtain confidence intervals. This option requires users to complete the conversion from "qualitative event to quantitative hypothesis" themselves. The conversion process is subjective and untraceable, and the correlation between the hypotheses usually requires users to manually specify the correlation coefficient matrix, which lacks a basis. Option 2: Pure Large Language Model Multi-Agent Simulation Scheme. Multiple large language model actors engage in multi-round game-like reasoning around an event, with the final decision-making actor summarizing and outputting a qualitative conclusion. Related research has been found in multi-agent simulation papers in engineering fields such as bridge maintenance and chemical process design, employing multi-agent role-playing combined with Monte Carlo methods to handle uncertainty. The simulation conclusions of this type of scheme are presented in natural language or a loose structure. The numerical judgments given by each actor do not have rigorously defined confidence intervals, and the correlation between judgments between actors is not systematically modeled. Furthermore, this type of scheme typically requires additional calls to the large language model to specifically infer the collaborative / correlation relationships between multiple actors (such as the method of parsing the large language model output into a weighted adjacency matrix, as seen in the literature), increasing system call costs and response latency. Option 3: Manual Correlation Matrix Specification. This option allows users to manually input correlation coefficient matrices between indicators within the numerical simulation engine to support multivariate joint sampling. This approach requires users to have a statistical background to manually estimate the correlation coefficients, and the estimated values ​​are often disconnected from the specific content of actual business events, relying more on the user's subjective experience than on information generated by the simulation process itself. The aforementioned existing technologies all have the following drawbacks: (1) There is a lack of automated quantitative bridging mechanism between the qualitative multi-role large language model simulation results and the deterministic numerical simulation engine. The two need to be transferred manually, and the conversion process is subjective and inefficient. (2) The numerical judgments of the output of the large language model roles lack uncertainty measures that are related to their own confidence level and can be used for quantitative simulation; (3) The correlation matrix between indicators depends on additional large language model calls or manual specification by the user. The former increases system cost, while the latter lacks objective basis. Summary of the Invention

[0003] Purpose of the invention The purpose of this invention is to provide a quantitative fusion method and system for simulation conclusions of multi-role large language models based on synthetic anchor nodes, so as to solve the technical problems in the prior art such as the separation between qualitative simulation conclusions and quantitative simulation engines, the lack of uncertainty measurement, and the high cost or lack of basis for correlation matrix estimation. Technical solution Core idea: Introduce a synthetic driving indicator that does not correspond to any actual business meaning in the base period indicator set to be simulated. Transform each qualitative indicator impact judgment generated by the multi-role large language model simulation into a transmission rule from the synthetic driving indicator to the target indicator. In this way, the qualitative judgment that could only be expressed as ordinary numerical assumptions (without uncertainty and correlation) is transformed into a ratio transmission rule with uncertainty and correlation that is natively supported by the deterministic numerical simulation engine. The same Monte Carlo engine is used to complete the quantization. The technical solution provided by this invention includes the following core mechanisms: Mechanism 1: Synthesis Anchor Bridge By introducing a constant synthetic driving index as the starting point for transmission, each qualitative judgment given by the large language model role is converted into the transmission ratio from the synthetic driving index to the target index. This transforms ordinary numerical assumptions that could not originally bear uncertainty and correlation into the native ratio transmission rules of the deterministic simulation engine. After the simulation is completed, the synthetic driving index is removed from the final result so that the user cannot see the internal implementation details. Mechanism 2: Confidence-Driven Uncertainty Mapping When a large language model role gives an indicator impact judgment, it also gives a confidence level label (such as high / medium / low). This invention converts the confidence level label into a relative standard deviation coefficient according to a preset deterministic mapping table. When the same indicator is judged by multiple roles, the most conservative (largest) standard deviation coefficient is taken, so that the confidence interval width of the quantitative simulation directly reflects the subjective confidence level of the role's judgment. Mechanism 3: Correlation inference based on co-occurrence relationships Instead of introducing additional large language model calls, this mechanism reuses the judgment data already generated by the large language model roles during simulation: if the same role gives a positive change to two indicators in the same judgment, it is considered to have cast a vote of "positive correlation"; if it gives a negative change, it is considered to have cast a vote of "negative correlation". The voting results of multiple roles are averaged to obtain the final correlation coefficient. This mechanism has zero additional call cost, and the correlation coefficient comes directly from the actual judgment content of the large language model roles, rather than being subjectively specified by the user. Beneficial effects (1) Eliminate manual conversion process: The conversion from qualitative events to quantitative hypotheses is fully automated, eliminating the need for users to manually complete the conversion from "qualitative judgment to numerical hypothesis". The conversion process is traceable and reproducible. (2) Uncertainty measurement is directly related to the confidence level of role judgment: Compared with the use of a uniform default uncertainty coefficient for all assumptions, this invention enables the confidence interval width of quantitative simulation to reflect the difference in confidence level of different role judgments, thereby improving the reference value of simulation results; (3) The cost of obtaining the correlation matrix is ​​zero: the correlation inference reuses the generated role judgment data, without generating additional large language model calls (compared to the approach of additionally calling large language models to infer collaborative / correlated relationships in Scheme 2, saving at least one model call delay and cost), and the correlation coefficient comes from the simulation process itself rather than the user's subjective specification (compared to Scheme 3), improving the objectivity and traceability of correlation estimation. Attached Figure Description Figure 1 This is a flowchart of the overall process of the synthetic anchor bridging method, showing the complete conversion process from the simulation output of a multi-role large language model to the quantized simulation results. Figure 2 A schematic diagram is constructed for the transmission rules of the synthetic anchor node, showing how the judgment of the influence of indicators is transformed into a transmission rule starting from the synthetic driving indicator. Figure 3 This is a flowchart of the deterministic mapping from confidence labels to standard deviation coefficients. Figure 4 This is a schematic diagram of the correlation matrix inference mechanism based on role-based judgment of co-occurrence relationships, showing how the judgment of the same role on multiple indicators is converted into correlation voting. Figure 5To quantify the overall architecture of the fusion system, a data flow relationship between various functional modules is shown. Detailed Implementation The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Example 1: Quantitative Bridging Scenario for Competitor Price Reduction Events A business decision support platform receives an external event input by a user: "Competitor X has reduced the price of its core product line by 20%". The platform knows that the company's base period gross profit margin is 30% (denoted as base period value 0.30). Step 1: Multi-role simulation generates qualitative judgments The platform calls on the finance role to perform simulation reasoning on the event. The financial role outputs the following indicator impact judgment: "Gross profit margin, change type: percentage change, change magnitude: -10, confidence level: high". Step 2: Synthesizing Anchor Point Bridging The platform adds a synthetic driving indicator (denoted as the anchor indicator, with a preset constant value of 1.0) to the base period indicator value set. Since the target indicator "gross profit margin" exists in the base period indicator value set, the quantitative target value is calculated according to the change type (percentage change): 0.30 × (1 + (-10) / 100) = 0.27. The transmission rule is constructed: anchor indicator → gross profit margin, ratio value = 0.27 / 1.0 = 0.27. Step 3: Confidence Mapping The confidence level label "High" is converted to a relative standard deviation coefficient of 0.08 according to a preset mapping table (mapping example: High → 0.08, Medium → 0.15, Low → 0.25). The ratio standard deviation of the transmission rule = |0.27| × 0.08 = 0.0216. Step 4: Deterministic Numerical Simulation The transmission rule, along with the assumption that the anchor index has a fixed value, is input into the Monte Carlo ratio transmission engine. The engine uses the constant value of the anchor index of 1.0 as the sampling mean benchmark, and samples 4000 times at a ratio value of 0.27 and a ratio standard deviation of 0.0216 to obtain the confidence interval estimate of the gross profit margin (e.g., the 90% confidence interval is approximately 0.256~0.284). Step 5: Result Removal and Restoration The simulation results exclude the data items corresponding to the anchor indicators. The final quantitative results output to the user only include the base period value, target value and confidence interval of the gross profit margin. The anchor indicators used internally are not visible to the user. If the existing technical solution 1 (where users manually complete the qualitative to quantitative conversion) is adopted, users need to determine "how many percentage points the gross profit margin will decrease due to a 20% price reduction and how wide the uncertainty range should be", which lacks a systematic basis. The solution of this invention is automatically completed by the role judgment of the large language model combined with deterministic mapping. The computational overhead added to the original multi-role simulation is negligible in terms of the total time consumed (the synthesis anchor bridging and confidence mapping are both lightweight deterministic calculations, not large language model calls). Example 2: Multi-indicator joint correlation inference scenario In the same event scenario, the financial role gives two judgments at the same time during the game correction round: "Gross profit margin, change range: -10, confidence level: high" and "Net profit margin, change range: -8, confidence level: high". Step 1: Record co-occurrence relationships According to the method of claim 7, the financial role assigns a negative indicator to the direction of change for "gross profit margin" (-10 is a negative number) and a negative indicator to the direction of change for "net profit margin" (-8 is a negative number). Since the two indicators are the same, it is recorded that the role casts the same vote in the "gross profit margin - net profit margin" indicator pair. Step 2: Calculation of Correlation Coefficient This indicator involves only the financial role, with 1 vote in the same direction and 0 votes in the opposite direction. The correlation direction coefficient is (1-0) / 1=1. Multiplying the preset fixed value of 0.5 by the correlation direction coefficient of 1 yields a final correlation coefficient of 0.5. Step 3: Joint Quantization Simulation The correlation coefficient of 0.5, along with the two transmission rules, is input into the multivariate normal distribution sampling to make the sampled values ​​of gross profit margin and net profit margin positively correlated—both are low in a pessimistic scenario and high in an optimistic scenario. This avoids sampling combinations that may occur due to independent sampling, such as "gross profit margin drops sharply but net profit margin is unaffected," which do not conform to business logic. If the existing technical solution three is adopted, the user needs to manually judge and input "what should the correlation coefficient between gross profit margin and net profit margin be set to"; the solution of this invention automatically derives a correlation coefficient of 0.5 from the two judgments actually given by the financial role, and this value directly reflects the actual judgment content of the role in this simulation, without generating additional large language model calling costs.

Claims

1. A quantitative bridging method for simulation conclusions of a multi-role large language model based on synthetic anchor nodes, used to convert the natural language form of the indicator influence judgment generated by the simulation of a multi-role large language model into a quantitative transmission structure that can be processed by a deterministic numerical simulation engine, characterized in that... Includes the following steps: (1) Obtain the set of indicator impact judgments from the simulation output of the multi-role large language model. Each judgment in the set of indicator impact judgments includes the target indicator identifier, change type, change magnitude and source role identifier. (2) Add a synthetic driving index to the set of base period index values ​​to be simulated. The value of the synthetic driving index is constant and does not correspond to any actual business meaning. (3) For each of the indicators in the set of indicator impact judgments and the target indicator identifier exists in the set of base period indicator values, calculate the quantitative target value of the target indicator according to its change type and change magnitude, and construct a transmission rule from the synthetic driving indicator to the target indicator. The ratio value of the transmission rule is equal to the quantitative target value divided by the preset constant of the synthetic driving indicator. (4) Mark the synthetic driving index as a hypothetical index with a determined value, and input all the transmission rules into the deterministic numerical simulation engine. The deterministic numerical simulation engine performs transmission calculations starting from the synthetic driving index to obtain the quantitative simulation results of all target indices. (5) After removing the data items corresponding to the synthetic driving index from the quantization simulation results, output the quantization simulation results after removal.

2. The method according to claim 1, characterized in that, The change type mentioned in step (3) includes two types: percentage change and absolute value change. When the change type is percentage change, the quantitative target value is equal to the base period value of the target indicator multiplied by (1 plus the percentage change). When the change type is absolute value change, the quantitative target value is equal to the base period value of the target indicator plus the change.

3. The method according to claim 1, characterized in that, The judgment that the target indicator identifier in the indicator impact judgment set described in step (1) does not exist in the base period indicator value set does not participate in the transmission rule construction in step (3) and is output separately as an unquantified impact.

4. The method according to claim 1, characterized in that, The deterministic numerical simulation engine is a ratio propagation simulation engine based on the Monte Carlo method; The transmission rule also includes a ratio standard deviation field; the transmission calculation in step (4) includes: taking the preset constant of the synthetic driving index as the mean benchmark, performing multivariate normal distribution sampling on all transmission rules according to the ratio value and ratio standard deviation of each transmission rule, and obtaining the confidence interval of each target index.

5. A method for determining the uncertainty of propagation based on role confidence labels in a large language model, applicable to the quantization bridging method described in claim 1, characterized in that... Includes the following steps: (1) Analyze the confidence label in the judgment of the influence of the indicator, wherein the confidence label is taken from a preset finite enumeration set; (2) Convert the confidence labels into corresponding relative standard deviation coefficients according to the preset deterministic mapping table from confidence labels to relative standard deviation coefficients; (3) When there are multiple judgments on the influence of the same target indicator from different roles, the largest relative standard deviation coefficient among the corresponding relative standard deviation coefficients shall be taken as the final relative standard deviation coefficient of the target indicator. (4) Multiply the absolute value of the quantitative target value of the target indicator by the final relative standard deviation coefficient to obtain the value of the ratio standard deviation field of the transmission rule in claim 1.

6. The method according to claim 5, characterized in that, The preset deterministic mapping table from confidence level labels to relative standard deviation coefficients includes at least three levels: the relative standard deviation coefficient corresponding to high confidence level is less than that corresponding to medium confidence level, and the relative standard deviation coefficient corresponding to medium confidence level is less than that corresponding to low confidence level.

7. A method for inferring the correlation matrix of indicators based on a multi-role large language model to determine co-occurrence relationships, used to provide correlation coefficients between indicators for the multivariate normal distribution sampling described in claim 4, characterized in that, Without relying on additional large language model calls, the following steps are included: (1) Traverse all the indicators output by each role during the simulation process to influence the judgment. For each role, construct a mapping from each target indicator involved by the role to its change direction identifier. The change direction identifier is determined by the positive or negative sign of the change magnitude of the judgment. (2) For any role involving two different target indicators, if the role gives the same direction of change for the two target indicators, it is counted as one vote in the same direction; if they are different, it is counted as one vote in the opposite direction. (3) For any target indicator pair that is involved by two or more roles at the same time, count the number of votes in the same direction and the number of votes against the target indicator pair given by all roles, calculate the sum of the number of votes in the same direction and the number of votes against the target indicator pair and divide it by the total number of votes to obtain the correlation direction coefficient of the target indicator pair. (4) Multiply the preset fixed amplitude value by the correlation direction coefficient to obtain the final correlation coefficient of the index pair; when the correlation direction coefficient is zero, the index pair does not output a correlation coefficient.

8. The method according to claim 7, characterized in that, In step (1), when the same role makes multiple judgments on the same target index in different rounds of simulation, the change direction indicator corresponding to the later round shall prevail.

9. A multi-role large language model simulation conclusion quantification and fusion system for implementing the methods of claims 1 to 8, characterized in that, include: - **Simulation Conclusion Acquisition Module**: Responsible for calling the multi-role large language model simulation and collecting the influence of the indicators output by each role on the judgment set; - **Synthetic Anchor Bridging Module**: Responsible for executing the construction of synthetic driving indicators, generation of transmission rules, and elimination and restoration of simulation results as described in claim 1; - **Confidence Mapping Module**: Responsible for performing the deterministic mapping from confidence labels to standard deviation coefficients as described in claim 5; - **Relevance Inference Module**: Responsible for performing the relevance matrix inference based on role judgment co-occurrence relationship as described in claim 7, without initiating additional large language model calls; - **Deterministic Numerical Simulation Engine Module**: Receives the transmission rules output by the Synthetic Anchor Bridging Module, the standard deviation output by the Confidence Mapping Module, and the correlation coefficient output by the Correlation Inference Module, performs multivariate normal distribution sampling, and outputs quantized simulation results.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

Images