Feature-Foundation Model Evolution: Toward Large-Model-Empowered Low-Latency Semantic Communication

To meet stringent requirements of high throughput in the 6G era, semantic communication (SC), a revolutionary framework that transmits semantic information rather than the original data, is expected to be a promising solution. However, existing SC frameworks struggle to accommodate data with new semantic features effectively and ensure minimal data processing latency. This is because achieving strong adaptability necessitates large-scale foundation models, such as large language models with robust semantic understanding, whereas low latency demands small models with simple structures. To address this issue, we propose a feature-foundation model evolution framework toward large model-empowered low-latency SC. In this framework, small-scale feature models and large-scale foundation models are deployed at edge nodes and cloud servers, respectively. The feature models perform real-time semantic encoding and decoding, and they evolve with the aid of foundation models when data with new semantic features occur, thus assuring low latency and high generalization ability simultaneously. We detail the architecture of the proposed collaborative framework and the related evolution protocol. Moreover, a case study on image transmission is presented to verify the effectiveness of the proposed framework. Furthermore, we summarize and discuss several future research directions based on the collaborative framework.