A new survey paper takes stock of federated prompt learning, an approach that trains large language models across decentralized devices without pooling raw data in one place.
The paper, posted to arXiv, reviews the field by tackling three questions: what motivates federated prompt learning and how it differs from standard federated learning or full-model federated fine-tuning; what trade-offs these methods show in performance, communication efficiency, computational overhead, scalability, personalization, and handling of heterogeneous clients; and what security, privacy, and robustness problems remain unsolved. The authors examine existing methods across pre-training, fine-tuning, and deployment, and summarize defense mechanisms proposed so far. It's a survey, not a new technique - the contribution is organizing a fragmented body of research, not shipping a model or benchmark result.
Federated learning has struggled to find a home with LLMs because full-model updates are expensive to move between devices and a central server. Tuning only prompts, rather than full model weights, is one way researchers have tried to make that exchange cheaper - though the survey itself doesn't quantify how much cheaper, or specify what kinds of clients are involved. That's the gap this paper says still needs closing: the trade-offs are catalogued, not resolved.
Surveys like this tend to arrive when a subfield has accumulated enough disconnected papers to need a map. Whether federated prompt learning becomes a standard tool or a footnote depends on work this paper doesn't do - actually benchmarking it against centralized fine-tuning at scale.