Abstract: In the traceability of power data links, large language models(LLMs) rely on fixed corpora for pre-training, making it difficult to adapt to rapidly changing complex relational networks. Their multi-step reasoning is often accompanied by illusions and omissions, especially in dynamic data and long chain dependency scenarios, making it difficult to maintain stable and accurate results. To address these issues, this paper proposes a dynamic iterative framework for the traceability of power data links based on LLMs. it incorporates dynamic adjustment of reasoning depth and introduces entity pruning during entity expansion, thereby reducing redundancy and invalid paths while ensuring reasoning efficiency and enhancing link completeness. The experimental results indicate that the proposed framework significantly improves the accuracy and reliability of link traceability, offering a new solution for knowledge-enhanced reasoning in complex power data environments.

Key words: traceability of power data links, large language models(LLMs), iterative reasoning, dynamic adjustment, entity pruning