Research on Intelligent Recommendation System of Audit Regulations Based on Large Language Model Driven by Multi-Task Fine-Tuning

The accuracy and efficiency of audit judgment heavily depend on the intelligence level of regulatory recommendation systems. Traditional methods face severe challenges due to issues such as dynamic updates of regulations and complex cross-domain associations. This paper proposes an intelligent regulatory recommendation framework that integrates multitask fine-tuning with retrieval-enhanced generation (RAG). By constructing a three-level task instruction set covering concept understanding, problem classification, and regulation matching (27,600 data points), and combining lightweight fine-tuning techniques (LoRA) to optimize the domain adaptability of the Qwen-7B model, experiments show that the fine-tuned model significantly outperforms baseline models in tasks such as audit issue summarization (ROUGE-L increased by 201%) and regulation classification (F1 reached 0.9977). After integrating RAG technology, the recommendation accuracy improves to 92.6%, and hallucination problems decrease by 64%. This study provides theoretical paradigms and practical tools for designing intelligent audit systems, promoting the transition of audit processes from “human-led” to “human-machine collaboration.”.