Toward Reliable Semantic Communication: Beyond Average Performance
For researchers in semantic communication, this paper provides a unified perspective on reliability beyond average performance, highlighting open challenges and future directions.
This paper addresses the reliability bottleneck in semantic communication under dynamic environments, where existing designs optimize for average performance but neglect lower-tail behavior. It reviews three reliability-oriented approaches, discusses their limitations, and proposes two solution directions: robust adaptive communication under imperfect CSI and joint source-channel-check coding with adaptive retransmission.
Semantic communication has emerged as a promising paradigm for improving transmission efficiency by conveying task-relevant semantics rather than raw data. Although recent studies have achieved notable gains in communication efficiency and average task performance, reliability remains a fundamental bottleneck in dynamic and uncertain environments. In particular, most existing designs are still optimized mainly for average-case behavior, while lower-tail performance under adverse transmission conditions remains insufficiently understood and inadequately protected. In this article, we present a unified perspective on reliable semantic communication beyond average performance. We first review three reliability-oriented design categories: channel-aware adaptation, robustness-oriented codec design, and hybrid automatic repeat request (HARQ)-based retransmission. We show that these approaches address reliability from complementary perspectives, but each still has inherent limitations. Motivated by these observations, we discuss two solution directions: robust adaptive semantic communication under imperfect CSI, and joint source-channel-check coding with adaptive retransmission for sample-level reliability enhancement. Finally, we outline several future research directions, including the joint design of robustness and retransmission, reliability metrics beyond averages, and compatibility with existing digital wireless networks.