Robustifying Token Communication Systems through Conformal Risk Control

Chenhao Wang, Zihan Chen, Tony Quek, Howard Hao Yang

Keywords: Token communications, generative semantic communications, large language models, conformal risk control, robust system, retransmission

Abstract: Token communication (TokCom) systems leverage the powerful semantic understanding of large language models (LLMs) to recover lost tokens, significantly enhancing communication efficiency and robustness against bursty losses. However, existing TokCom frameworks rely solely on LLM completion and completely fail to validate the correctness of the recovered tokens, leading to unacceptable error risks in critical tasks. To address this gap, we propose the \textit{first} retransmission mechanism for TokCom systems. Unlike conventional retransmission mechanisms that unconditionally retransmit error tokens, our mechanism is based on Conformal Risk Control to learn a decision threshold with theoretical guarantees, enabling selective retransmission without any assumptions about the internal structure of the LLMs. Extensive simulation results on text and image tasks show that our method significantly improves communication robustness compared with conventional TokCom systems that rely solely on LLM completion, while requiring significantly fewer retransmission requests than traditional protocols.

Submission Number: 16