CN-122021946-A - Large language model decision method integrating causal reasoning and uncertainty quantification

CN122021946ACN 122021946 ACN122021946 ACN 122021946ACN-122021946-A

Abstract

The invention discloses a large-scale language model decision method and system integrating causal reasoning and uncertainty quantification, and aims to solve the problems of unreliable, unexplainable and lack of auditability of automatic decisions in the high-risk field. The method comprises the steps of firstly using a large language model as a state inspirer, extracting state variables from natural language situations and modeling random uncertainty of the state variables, then constructing a probability causal graph comprising causal dependency relations and confidence of the cognitive uncertainty, converting the probability causal graph into a directed acyclic graph through a confidence-driven greedy strategy, carrying out value assessment on the cognitive uncertainty as an attenuation factor by adopting a confidence-weighted causal propagation algorithm, and finally deriving an optimal strategy and generating auditable causal evidence based on a maximum expected utility criterion. Experiments show that the invention obviously improves the decision accuracy in the agricultural and financial decision tasks, and realizes more robust, transparent and credible automatic decision based on model planning.

Inventors

ZHOU HOUKUI
LI XINZHAN
LI CHENGXUAN
NIE YUAN
GUO SHUTONG

Assignees

浙江农林大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (8)

1. A large language model decision method integrating causal reasoning and uncertainty quantification is characterized by comprising the following steps: S1, a state reasoning step of giving a decision context of natural language description And a preset executable action set From the decision context using a large language model Extracting state variable set from For each state variable Evaluate its target result Is the expected effect of (a) Modeling random uncertainty of the environment as probability distribution to obtain formalized variables Wherein A natural language tag that is a state variable, For the expected effect of the variable on the target result, As an inherent fluctuation of the effect, The result dimension for the variable influence; S2, constructing a causal graph, namely utilizing a large language model to construct a state variable set Each pair of variables in (a) Make a pair-wise causal determination to determine if a direct causal relationship exists And quantifying the reliability of reasoning while deducing the causal relationship, generating confidence label Constructing a set of containing nodes Edge set Probability distribution Confidence level labeling Probability causal graph of (2) Wherein each causal edge Is formally expressed as , Reasoning basis for supporting causal inference; s3, loop elimination, namely detecting and removing edges with lowest confidence in a loop by adopting a confidence driving greedy strategy until the graph structure becomes a directed acyclic graph ; S4, a causal weight calculation step of a directed acyclic graph Performing topological sorting, and recursively calculating causal weights of $v$ of each node according to topological reverse order : A, when In the case of a leaf node, the node is, ; B, when In the case of a non-leaf node, , Wherein the method comprises the steps of Representing directed acyclic graphs Is provided with a set of leaf nodes in the tree, Is a node Is defined by a set of sub-nodes of the set, Is a side Confidence of (2); S5, expected utility calculating step, utilizing causal weight To the expected influence Performing weighted aggregation and calculating the result Each dimension is provided with Expected change in weighting of (2) And calculate each action Is of desired utility of (3) ; S6, a decision generation step of selecting an action capable of bringing the highest expected utility as an optimal strategy And generating a containing causal graph Is provided for the auditability report of (a).
2. The method according to claim 1, wherein in said step S1, said probability distribution is a normal distribution Modeling state variables Random uncertainty of (1), wherein Representing the expected effect of the variable on the target result, Indicating the inherent volatility of this effect.
3. The method according to claim 1, wherein in said step S2, said confidence score is Extracting variables for quantifying large language model Or infer causal relationships The cognitive uncertainty in the time of day, The closer to 1 the value of (c) indicates a higher confidence that the causal relationship exists.
4. The method of claim 1, wherein in step S3, the confidence-driven greedy strategy is based on the principle that among the set of edges that make up the loop, the edge with the lowest confidence represents the reasoning error or the weakest causal link of the large language model, thus iteratively detecting and removing the loop Lowest edges until the graph structure becomes a directed acyclic graph 。
5. The method according to claim 1, wherein in said step S5, the result is Each dimension is provided with Expected change in weighting of (2) The calculation formula of (2) is as follows: Wherein the method comprises the steps of Is to influence the dimension of the result Variable set of (2), the aggregate value All relevant variables' contributions to this dimension are included and subjected to a double calibration of causal weights and confidence.
6. The method according to claim 1, wherein in said step S5, each action is calculated Is of desired utility of (3) The formula of (2) is: Wherein the method comprises the steps of As a utility function, where As a function of the utility, As a basis for the utility of the present invention, A change is expected for the weighting of each result dimension.
7. The method according to claim 1, wherein in said step S6, said auditability report comprises all actionable actions Ranking of desired utility for optimal action The highest weighted core cause and effect path, and the reasoning basis for each decision step.
8. The method of claim 1, wherein the decision context Including observable state variables And unobserved variables affecting decision results The action is Aimed at influencing these variables, ultimately resulting in a resultant state 。

Description

Large language model decision method integrating causal reasoning and uncertainty quantification Technical Field The invention relates to the technical field of artificial intelligence, in particular to a large-scale language model decision-making method integrating causal reasoning and uncertainty quantification, and especially relates to a technical scheme for making a reliable decision in an uncertainty environment. Background In fields with extremely high security requirements, such as finance, healthcare, and critical infrastructure management, the ability of automated systems to provide reliable and interpretable decisions has become a core technical challenge. The efficient decision-making does not rely solely on pattern matching, but rather requires a rigorous structuring process by first identifying all key variables that affect the outcome of the action, then accurately modeling the causal interactions between these factors, and on that basis systematically evaluating the expected utility of the various behaviors in an uncertainty environment. The structured reasoning based on the causal logic is not only a basic stone for improving the accuracy of the decision, but also a key guarantee which ensures the transparency, auditability and responsibility of the decision process. While Large Language Models (LLMs) are increasingly becoming ideal choices for complex decision support tools with strong reasoning capabilities, existing decision paradigms still face severe technical limitations. The core problem is that the mainstream end-to-end generation method often lacks explicit modeling of the underlying causal structure, and is difficult to provide a robust decision strategy in a complex scene. While decision theory and utility theory have already provided formalized frameworks for complex choices based on the "maximum expected utility" criterion, the current LLM paradigm is still deficient in the explicit representation of decision space and simulation of causal interactions, resulting in difficult efficient quantification and management of its decision basis. In addition, the existing LLM decision process generally has the problems of logic fragmentation and causal chain weakness, so that decision basis is difficult to trace back. When the uncertainty is processed, the model often shows obvious index processing inconsistency, and random uncertainty of the environment and cognitive uncertainty of the model can not be effectively distinguished and quantified. More controversially, the interpretation of these model formations is typically constructed "posterior", i.e., the logic is matched after the results are given, rather than naturally occurring from the reasoning process. This lack of endogenous audit mechanisms severely undermines the public confidence of decisions in areas where high resolution is required. Therefore, a new technical solution is needed to introduce explicit causal modeling depth into the decision loop of LLM. By constructing a structured causal reasoning framework, unstructured natural language situation is converted into a computable and traceable causal model, so that reliable and auditable automatic decision in a high-risk scene can be truly realized, and the huge gap of the existing generation formula in the aspect of reliability is filled. Disclosure of Invention The invention aims to provide a large-scale language model decision method integrating causal reasoning and uncertainty quantification, which aims to solve the problems in the prior art. The invention provides a large-scale language model decision method integrating causal reasoning and uncertainty quantification, which comprises the following steps: (1) A state reasoning step of giving a decision context of natural language description And a preset executable action setFrom the decision context using a large language modelExtracting state variable set fromFor each state variableEvaluate its target resultIs the expected effect of (a)Modeling random uncertainty of the environment as probability distribution to obtain formalized variablesWhereinA natural language tag that is a state variable,For the expected effect of the variable on the target result,As an inherent fluctuation of the effect,For the result dimension of the variable influence, this step converts unstructured text into a set of computable elementsThese elements cover influencing the decision resultBy modeling the randomness (i.e., random uncertainty) of the environment as a probability distribution; (2) Causal graph construction step, namely, a large language model is utilized to construct a state variable set Each pair of variables in (a)Make a pair-wise causal determination to determine if a direct causal relationship existsAnd quantifying the reliability of reasoning while deducing the causal relationship, generating confidence labelConstructing a set of containing nodesEdge setProbability distributionConfidence level labelingProbability causal grap