SANet: A Semantic-Aware Agentic AI Networking Framework for Cross-Layer Optimization in 6G
IEEE Transactions on Mobile Computing · 2026-09 · 25 (9) · 14849-14866
DOI: 10.1109/tmc.2026.3691804 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Mobile Computing, 2026-09. Verfasst von Yong Xiao und 7 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 36 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Mobile Information Network National Science and Technology Key Project, Hubei Natural Science Foundation Innovation Research Group Program.
Abstract (Originalsprache)
Agentic AI networking (AgentNet) is a novel AI-native networking paradigm in which a large number of specialized AI agents collaborate to perform autonomous decisions, dynamic adaptations, and complex missions. Despite its potential to enable real-time autonomous networking, AgentNet is still in the early stages of development and lacks an effective framework for goal discovery, self-orchestration, and task assignment. To bridge this gap, we propose SANet, a novel semantic-aware AgentNet architecture for wireless networks that infers users' semantic goals and automatically assigns multi-layer agents to fulfill them. Recognizing that AgentNet is a decentralized framework and different agents often have conflicting objectives, we formulate decentralized optimization of SANet as a multi-agent, multi-objective problem focused on finding Pareto-optimal solutions. We introduce three novel evaluation metrics: optimization error, generalization error, and conflicting error. To optimize these metrics, we develop a Model Partition and Sharing (MoPS) framework that partitions large models into shared and agent-specific components based on agents' local resource constraints. We propose static- and dynamic-weighting optimization algorithms, as well as a bandwidth-adaptive compression framework that enables agents to dynamically compress intermediate embeddings based on resource constraints. We derive theoretical bounds and prove the fundamental three-way tradeoff among these metrics. Finally, we validate our findings using an open-source RAN and core network hardware prototype employing Transformer model-based agents. Experiments show our MoPS framework delivers up to a 14.61% performance gain while requiring only 44.37% of the inference FLOPs compared to state-of-the-art algorithms. Additionally, our dynamic-weighting algorithm reduces training errors caused by conflicting objectives by up to 83.81%.
Förderung
- National Natural Science Foundation of China
- Mobile Information Network National Science and Technology Key Project
- Hubei Natural Science Foundation Innovation Research Group Program
Zitieren
APA
Xiao, Y., Li, X., Zhou, H., Li, Y., Gao, Y., Shi, G., Zhang, P., & Krunz, M. (2026). SANet: A Semantic-Aware Agentic AI Networking Framework for Cross-Layer Optimization in 6G. IEEE Transactions on Mobile Computing, 25(9), 14849-14866. https://doi.org/10.1109/tmc.2026.3691804
Vancouver
Xiao Y, Li X, Zhou H, Li Y, Gao Y, Shi G, et al. SANet: A Semantic-Aware Agentic AI Networking Framework for Cross-Layer Optimization in 6G. IEEE Transactions on Mobile Computing. 2026;25(9):14849-14866. doi:10.1109/tmc.2026.3691804
BibTeX
@article{Xiao2026,
author = {Xiao, Yong and Li, Xubo and Zhou, Haoran and Li, Yingyu and Gao, Yayu and Shi, Guangming and Zhang, Ping and Krunz, Marwan},
title = {SANet: A Semantic-Aware Agentic AI Networking Framework for Cross-Layer Optimization in 6G},
journal = {IEEE Transactions on Mobile Computing},
year = {2026},
volume = {25},
number = {9},
pages = {14849--14866},
doi = {10.1109/tmc.2026.3691804}
}
EUVIMED