Personalized Federated Prompt Learning for Vision-Language Models: A Survey

Federated Prompt Learning (FPL) has emerged as an efficient paradigm for adapting large Vision-Language Models (VLMs) in decentralized environments. By optimizing lightweight prompt parameters instead of full model weights, FPL reduces communication and computation costs while preserving the generalization ability of foundation models. However, data heterogeneity across clients makes personalization a core issue in practical federated settings. Diverse data distributions, task preferences, and sample scales often limit the effectiveness of a single global prompt. This survey provides a systematic review of recent advances in personalized FPL. We first introduce the basic formulation of FPL and summarize common prompt representation, including textual, visual, and multimodal prompts. We then classified the existing personalized methods from the perspectives of granularity and mechanism, covering decoupled prompts, prompt generation, and prompt selection. We further provide a comparative analysis of these mechanisms and carry out an empirical study on common methods. Furthermore, we discuss key challenges related to robustness under data heterogeneity, personalization under black-box constraints, and privacy risks in FPL. By clarifying the trade-offs of personalized FPL, this survey aims to provide a structured reference for researchers working on personalized federated adaptation of VLMs.