MambaMTSC: A Unified State-Space Modeling Framework for Multimodal and Multi-Task Semantic Communication
IEEE Transactions on Communications · 2026 · 74 · 12151-12167
DOI: 10.1109/tcomm.2026.3716996 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Communications, 2026. Verfasst von Puning Zhang und 4 weiteren Autor:innen. Die Arbeit nennt 59 Literaturangaben. Gefördert von: National Key R&D Program of China, Chongqing Municipal Key Project for Technological Innovation and Application Development, Science and Technology Research Program of Chongqing Municipal Education Commission.
Abstract (Originalsprache)
Existing semantic communication systems are typically designed for a single modality or a single task. Recent multimodal multi-task semantic communication frameworks are predominantly built on Transformer architectures, which incur quadratic computational complexity, limited efficiency for long-sequence modeling, and poor adaptability to time-varying channel conditions. To address these challenges, this paper proposes MambaMTSC, a unified framework for efficient and robust semantic transmission. To the best of our knowledge, MambaMTSC is the first unified multimodal multi-task semantic communication framework built upon a selective state-space model (SSM). First, a unified Mamba-based semantic encoder is developed to model long multimodal sequences with linear complexity while disentangling task-agnostic public features and task-specific private features through task control parameters. Second, a separated Mamba fusion module (SMFM) is proposed to construct modality pseudo-sequences and perform bidirectional state-space recursion, thereby enabling dynamic cross-modal aggregation of both public and private feature streams. Third, a channel-state-aware conditional transmission mechanism is designed by explicitly incorporating the signal-to-noise ratio (SNR) as a conditioning variable, allowing adaptive semantic compression at the transmitter and two-stage noise-robust semantic recovery at the receiver, thereby balancing compression efficiency and robustness under varying channel conditions. Experimental results demonstrate that MambaMTSC significantly outperforms state-of-the-art baselines in multi-task performance, while substantially reducing the overhead required for long-sequence processing. Additionally, it exhibits remarkable robustness, particularly in low-SNR regimes.
Förderung
- National Key R&D Program of China
- Chongqing Municipal Key Project for Technological Innovation and Application Development
- Science and Technology Research Program of Chongqing Municipal Education Commission
- Natural Science Foundation of Chongqing Municipality
- National Natural Science Foundation of China
- China Postdoctoral Science Foundation
- Chongqing Postdoctoral Research Program Special
Zitieren
APA
Zhang, P., Sun, J., Wu, C., Sun, M., & Yang, Z. (2026). MambaMTSC: A Unified State-Space Modeling Framework for Multimodal and Multi-Task Semantic Communication. IEEE Transactions on Communications, 74, 12151-12167. https://doi.org/10.1109/tcomm.2026.3716996
Vancouver
Zhang P, Sun J, Wu C, Sun M, Yang Z. MambaMTSC: A Unified State-Space Modeling Framework for Multimodal and Multi-Task Semantic Communication. IEEE Transactions on Communications. 2026;74:12151-12167. doi:10.1109/tcomm.2026.3716996
BibTeX
@article{Zhang2026,
author = {Zhang, Puning and Sun, Jie and Wu, Chao and Sun, Meiyu and Yang, Zhigang},
title = {MambaMTSC: A Unified State-Space Modeling Framework for Multimodal and Multi-Task Semantic Communication},
journal = {IEEE Transactions on Communications},
year = {2026},
volume = {74},
pages = {12151--12167},
doi = {10.1109/tcomm.2026.3716996}
}
EUVIMED