Towards a Universal Semantic Basis: A Survey on the Fundamental Units of Semantic Communication in the LLM Era

As communication technology enters the 6G era, traditional "bit-based" syntax-level (Level A) communication is hitting physical bottlenecks from the Shannon limit. In recent years, the rise of Large Language Models (LLMs), especially technologies aimed at distributed deployment and prior knowledge sharing, has made LLMs a strong candidate for shifting human communication to "intent and semantics" (Level B) transmission. Moving semantic communication toward engineering practice and standardisation requires a unit equivalent to "bits" in traditional systems. However, most current studies on LLM collaborative communication focus on edge-cloud synchronisation of intermediate state tensors like the KV Cache. This article introduces the structural characteristics of the KV Cache and shows why it cannot serve as the minimum measurement unit for semantic communication. Meanwhile, building on existing intelligent agent collaboration technologies, this article derives the basic properties the fundamental unit of pan-semantic communication must possess. The analysis shows that intermediate tensors such as the KV Cache are constrained by strong model coupling and dimensional expansion; they serve only as temporary computational residues rather than true semantic carriers. Therefore, this paper concludes that a valid minimal unit of semantic communication must be model-agnostic and topologically invariant. Defining such a unified, intent-based unit helps communication systems move from rigid state synchronisation to flexible semantic routing in future 6G networks.