FedPattern: A Communication Efficient Federated Pattern Learning Agent Framework for Cyber-Physical-Social Systems

Zhixiang Zhang, Zhicheng Bao, Dehai Zhao, Qinghua Lu, Guodong Long, Jing Jiang, Liming ZhuIEEE Transactions on Computational Social Systems 2025 · DOI

Abstract

Large language models (LLMs) begin to be widely used in cyber-physical-social systems (CCPS). Existing federated learning (FL) solutions to fine-tune LLMs usually incur significant communication overhead. Strategic activation of LLMs’ inherent knowledge reservoirs and logical reasoning capacities through task-specific prompts is promising to address this problem, with the possibility of improving model accuracy. In this article, we design a communication efficient federated Pattern learning agent framework for CCPS. First, we propose an agent-based gradient-free knowledge learning method, where domain task rules are extracted into explicit knowledge representations, as called Pattern in this article. The Pattern fills domain-specific knowledge gaps in LLMs while guiding their reasoning processes within specialized domains. Second, to achieve distributively collaborative training, we design a Primary-Secondary Pattern Aggregation Algorithm to replace existing FL parameter fusion with Pattern fusion, achieving a significant reduction in communication overhead. Comprehensive evaluations demonstrate that this framework can effectively extract useful patterns, achieving gradient-free adaptation of foundation models to downstream tasks. Extensive experiments show that, through FL, the proposed FedPattern approach has significant improvements over existing solutions, which can improve accuracy by 20.34%, while achieving 99.99% communication overhead reduction compared to that of the prompt-based FL baselines.
BibTeX
@article{zhang2025fedpattern,
  author = {Zhixiang Zhang and Zhicheng Bao and Dehai Zhao and Qinghua Lu and Guodong Long and Jing Jiang and Liming Zhu},
  title = {{FedPattern: A Communication Efficient Federated Pattern Learning Agent Framework for Cyber-Physical-Social Systems}},
  journal = {IEEE Transactions on Computational Social Systems},
  year = {2025},
  doi = {10.1109/tcss.2025.3634841},
}

Loading related content...

Figure