Cognitive Agentic Semantic Wireless Networks for AI-Native 6G Communication Systems

Abstract

Abstract—The rise of Internet of Things (IoT) devices and the advent of 6th generation (6G) wireless networks necessitate a fundamental revision of communication paradigms. The explosive growth of heterogeneous device data, stringent ultra-reliable low-latency communication (URLLC) requirements, and severe energy constraints at the edge terminals make the traditional bit-pipe transmission paradigms unsustainable. In this paper, we propose the Cognitive Agentic Semantic Wireless Network (CASWN), a novel five-layer AI-native network architecture that converts the information transmission unit from raw bits to semantic representations of relevant tasks. CASWN contains three independent cognitive agents, a bandwidth agent, an energy agent, and a latency agent, which together control the semantic feature transmission volume according to the real-time network conditions. We use Principal Component Analysis (PCA) as the semantic encoder to compress the 561-dimensional smartphone inertial sensor signals into task-essential feature subspaces. The compressed semantic representation is then used for human activity recognition (HAR) with a Random Forest classifier deployed at the edge intelligence layer. Comprehensive experimental evaluation on the UCI Human Activity Recognition (HAR) dataset demonstrates that CASWN achieves a peak classification accuracy of 99.15% and a weighted F1-score of 99.14%, while achieving communication overhead reductions of up to 93.05% in high-compression scenarios. We demonstrate that CASWN yields better per-bit semantic value than baseline and prior-art methods by using the intelligence efficiency metric defined as F1-score divided by the transmitted feature ratio. The proposed architecture is benchmarked against seven recent state-of-the-art methods (2021-2025), and consistently achieves higher accuracy than these methods by 1.35%-7.95%. CASWN: A principled foundation for cognitive semantic communications in AI-native 6G IoT ecosystems.