Federated Prompt Learning Survey Maps Privacy-Preserving Path for LLMs

A research team led by Qinglin Yang has released a comprehensive survey on federated prompt learning (FPL), a technique that marries federated learning with large language models (LLMs) to enable privacy-preserving training and fine-tuning. The paper, published on arXiv (2608.13844), systematically analyzes how FPL works, why it differs from conventional federated learning and full-model fine-tuning, and what trade-offs it makes in performance, communication efficiency, computational overhead, scalability, and personalization.

The survey organizes existing FPL methods across the entire machine learning lifecycle, from pre-training to fine-tuning and practical deployment, and devotes special attention to security, privacy, and robustness. It catalogs defense mechanisms and outlines key open challenges, including vulnerability to adversarial attacks, data heterogeneity, and the difficulty of balancing personalization with global model coherence. By answering three core research questions—motivations and enabling technologies, trade-offs across approaches, and remaining challenges—the authors provide a structured roadmap for researchers and practitioners. This positions FPL as a critical stepping stone toward decentralized, privacy-compliant LLM systems that can be deployed across industries without centralizing sensitive data.