Large Semantic Agents for Wireless Image Transmission
Wei Yuan, Jinke Ren, Rui Sun, Yifei Han, Shuguang Robert Cui · 2025
Semantic communication (SemCom) is recognized as a promising technology to improve communication efficiency by transmitting the meaning of data rather than raw bits. However, existing SemCom systems are typically designed for specific tasks with limited generalization and reasoning abilities. To address these issues, this paper considers a point-to-point image transmission system and proposes two large language model (LLM)-based semantic agents, namely LSAs, which serve as the semantic codecs at the transmitter and receiver, respectively. Each LSA comprises three key components, including a multi-source prompt module, a pre-trained LLM, and an auxiliary tool. The multi-source prompt module first transforms the input data and environmental information into token-like features that can be processed by the pre-trained LLM. Then, the LLM performs semantic understanding at the transmitter or content regeneration at the receiver, thereby producing semantic embeddings or generating images. The auxiliary tool consists of application programming interfaces and databases, assisting the LLM in semantic understanding and image optimization. Experimental results demonstrate that the proposed LSA-based SemCom scheme achieves superior image transmission performance than four baseline schemes.