This paper studies task-oriented communication for cooperative edge inference in wireless sensor networks, where multiple edge devices transmit compact feature representations to a central edge server for joint decision-making. Recently, cooperative edge inference significantly improved the device-server system performance, however at the expense of increased communication overhead. To this end, an adaptive variational information bottleneck (A-VIB) framework that enables edge devices to extract and transmit task-relevant features under dynamic channel conditions is proposed. The framework jointly optimizes feature relevance, communication efficiency, and inter-device coordination through stochastic adaptive encoding and graph-based mutual-information suppression. To further reduce bandwidth consumption, a graph-coordinated feature decorrelation strategy is developed, enabling each device to adjust its feature dimensionality and minimize transmitting redundant information. It employs a local redundancy suppressor on the edge device to discard task-irrelevant features during extraction. We also adopt the information bottleneck principle and propose a dynamic rate control network module (RateNet) to accommodate the varying networking conditions. Extensive experiments on multi-view image classification and multi-view object recognition tasks demonstrate the effectiveness of A-VIB, achieving up to a 58.6% reduction in communication cost and up to a 13.2% improvement in inference accuracy compared with state-of-the-art semantic communication and distributed inference schemes under varying SNR and bandwidth-constrained wireless environments. In addition, A-VIB exhibits improved robustness to channel interference by 3.29%.

The growing demand for intelligent and adaptive services is accelerating the transition from conventional communication systems to semantic communication. Task-Oriented Semantic Communication (ToSC) has emerged as an efficient framework that reduces communication cost by transmitting only task-relevant information for downstream processing. This survey reviews the theoretical foundations, system architectures, and enabling technologies of ToSC, providing a taxonomy of design strategies spanning collaborative edge inference, end-to-end optimization, and multi-agent cooperation. Key technologies, including deep learning-based semantic representation and adaptive joint source-channel coding, are discussed. Device heterogeneity and partial observability are examined alongside recent tasks' performance evaluations, such as classification and recognition. The survey also discusses the integration of large-scale multimodal foundation models into semantic encoders for zero-shot task adaptation, continual learning frameworks for on-device semantic processing, cooperative multi-agent reasoning with interpretable information exchange, and cross-layer co-design within AI-native 6G infrastructures, leveraging reconfigurable intelligent surfaces and digital twins. By consolidating existing knowledge and identifying key design trade-offs in latency, accuracy, and bandwidth efficiency, this work aims to serve as both a reference and a roadmap for advancing ToSC toward reliable and resource-efficient communication in 6G and beyond.