LLMs 监督的因果威胁评估模型

To address the challenge of building a threat assessment network that makes data distribution and expert knowledge compatible while preserving causal interpretability, L-CTA, an LLMs-supervised causal threat assessment model is proposed. The expert roles of LLMs are utilized in simulating the traditional modeling, the prompts are fused to construct the causal diagram of threat factors and decision variables. The parameter learning is performed based on simulated data, the inference accuracy on the verification set is verified by k-fold cross-validation and taken as the fitness of the causal diagram, the rational evolutionary operators are introduced, the more optimal diagram structure is found by iterative searching. The experiments show that the causal threat assessment network constructed by L-CTA achieves 11.2% and 28.9% higher accuracy than those of naive model designed by experts and data-driven models in out-of-distribution inference tasks, its efficiency, robustness, and generalization in threat assessment of modeling of L-CTA are verified.