Context-Aware Semantic Communication for the Wireless Networks

TL;DR

CaSemCom tackles the challenge of transmitting only task-relevant semantic content over dynamic wireless links, addressing inefficiencies of static SemCom pipelines. It introduces an LLM-based gating mechanism combined with a Mixture of Experts to dynamically select input content and activate modality-specific encoders, optimizing semantic fidelity under bandwidth and latency constraints. The paper demonstrates a multimodal, multi-user case study showing faster convergence, higher $SSIM$ and $ImageReward$, and reduced retransmission overhead under Rayleigh fading with $SNR$ in $[-13,30]$ dB. The work advances context-aware SemCom by unifying linguistic, wireless, and semantic cues and enables scalable, reusable semantic encoders for future wireless systems. The practical impact lies in supporting real-time AR/VR, autonomous driving, and Metaverse services with lower bandwidth while preserving meaning.

Abstract

In next-generation wireless networks, supporting real-time applications such as augmented reality, autonomous driving, and immersive Metaverse services demands stringent constraints on bandwidth, latency, and reliability. Existing semantic communication (SemCom) approaches typically rely on static models, overlooking dynamic conditions and contextual cues vital for efficient transmission. To address these challenges, we propose CaSemCom, a context-aware SemCom framework that leverages a Large Language Model (LLM)-based gating mechanism and a Mixture of Experts (MoE) architecture to adaptively select and encode only high-impact semantic features across multiple data modalities. Our multimodal, multi-user case study demonstrates that CaSemCom significantly improves reconstructed image fidelity while reducing bandwidth usage, outperforming single-agent deep reinforcement learning (DRL) methods and traditional baselines in convergence speed, semantic accuracy, and retransmission overhead.

Figures (5)

• Figure 1: A unified illustration of linguistic communication, SemCom, and wireless transmission, each governed by internal and external contexts. Linguistic communication emphasizes how cognitive abilities, knowledge bases, and situational cues shape the exchange of human language. Wireless communication frames the physical layer, where device constraints, channel fading, and interference further refine which data can be transmitted in real-time. SemCom captures meaning extraction and representation, highlighting how semantic encoders and task objectives drive the transformation of raw data into high-level features.
• Figure 2: Architecture of the proposed LLM-based CaSemCom framework. At the transmitter, the system determines which parts of the data are semantically important and selects the best expert(s) to encode those parts. The receiver uses the same gating information to ensure accurate semantic reconstruction.
• Figure 3: An overview of the LLM-based context gate and its fallback mechanism. Task context ($C_{T}$) and communication context ($C_{c}$) serve as inputs to the LLM, which outputs content-selection and expert-selection decisions. These decisions, alongside the context parameters, are stored in an experience pool as state-action pairs. Over time, a DRL agent is trained on this data and can serve as a fallback decision-maker when LLM inference is unavailable or infeasible.
• Figure 4: Evolution of the average reconstructed image quality, measured by a combined Image Reward and SSIM metric, over 250 training steps. CaSemCom with LLM converges most rapidly and achieves the highest final reward, whereas CaSemCom with DRL fallback converges more gradually but still significantly outperforms the pure DQN, greedy, and random baselines. The DeepSC baseline, despite its learning-based strategy, performs notably worse than CaSemCom with LLM and CaSemCom with DRL fallback, indicating the effectiveness of context-aware gating mechanisms in dynamic semantic communication environments.
• Figure 5: Retransmission overhead (log scale) for six methods (CaSemCom with LLM, CaSemCom with DRL fallback, Pure DQN, DeepSC, greedy, and random selection) under different semantic fidelity requirements (0.3, 0.6, and 0.9). CaSemCom with LLM consistently achieves the lowest retransmission overhead, highlighting the advantage of context-aware gating in bandwidth-constrained scenarios.