To address the challenge of characterizing the knowledge states of large-scale language models (LLMs), this paper introduces a five-class knowledge state taxonomy based on the consistency and accuracy of LLM knowledge modes. We propose a hierarchical statistical testing framework, KScope, to identify these knowledge states and apply it to four datasets and nine LLMs. We demonstrate that (1) supporting context reduces the knowledge gap between models; (2) contextual features related to difficulty, relevance, and familiarity drive knowledge update success; (3) LLMs exhibit similar feature preferences when partially accurate or conflicting but differ significantly when consistently incorrect; and (4) limited contextual summarization and enhanced reliability through feature analysis enhance update efficiency and generalize across a wide range of LLMs.
