Towards Artificial General or Personalized Intelligence? A Survey on Foundation Models for Personalized Federated Intelligence

TL;DR

The paper addresses the challenge of tailoring large foundation models to individual users while preserving privacy and efficiency by proposing artificial personalized intelligence (API) and the personalized federated intelligence (PFI) framework. It surveys foundational topics in federated learning and foundation models, then outlines a design space that combines retrieval-augmented generation (RAG) with federated optimization to enable end-user personalization at the edge. Key contributions include a taxonomy of FM–FL interactions, concrete RAG-based personalization strategies (client-side prompts, local indexing, retriever adaptation, server-side task matching and clustering), and efficiency and trust-focused techniques (PEFT, pruning, quantization, RLHF/RLAIF, XAI, and robust defenses), along with challenges and future directions such as Meta-PFI, quantum-enabled PFI, and sustainable green PFI. The work underscores API as a practical, privacy-preserving complement to AGI, with significant implications for deploying customizable AI at scale on resource-constrained devices.

Abstract

The rise of large language models (LLMs), such as ChatGPT, DeepSeek, and Grok-3, has reshaped the artificial intelligence landscape. As prominent examples of foundational models (FMs) built on LLMs, these models exhibit remarkable capabilities in generating human-like content, bringing us closer to achieving artificial general intelligence (AGI). However, their large-scale nature, sensitivity to privacy concerns, and substantial computational demands present significant challenges to personalized customization for end users. To bridge this gap, this paper presents the vision of artificial personalized intelligence (API), focusing on adapting these powerful models to meet the specific needs and preferences of users while maintaining privacy and efficiency. Specifically, this paper proposes personalized federated intelligence (PFI), which integrates the privacy-preserving advantages of federated learning (FL) with the zero-shot generalization capabilities of FMs, enabling personalized, efficient, and privacy-protective deployment at the edge. We first review recent advances in both FL and FMs, and discuss the potential of leveraging FMs to enhance federated systems. We then present the key motivations behind realizing PFI and explore promising opportunities in this space, including efficient PFI, trustworthy PFI, and PFI empowered by retrieval-augmented generation (RAG). Finally, we outline key challenges and future research directions for deploying FM-powered FL systems at the edge with improved personalization, computational efficiency, and privacy guarantees. Overall, this survey aims to lay the groundwork for the development of API as a complement to AGI, with a particular focus on PFI as a key enabling technique.

Figures (14)

• Figure 1: The structure and scope of the survey.
• Figure 2: Illustration of cross-device and cross-silo federated learning scenarios.
• Figure 3: Illustration of horizontal federated learning, vertical federated learning, and federated transfer learning liu2024vertical.
• Figure 4: A taxonomy of techniques for enabling retrieval-augmented PFI. The design is organized into two complementary pillars: (1) Client-side personalization, encompassing methods such as prompt conditioning, local knowledge indexing, and retriever adaptation using LoRA or adapters; and (2) Server-side knowledge-augmented optimization, which includes retrieval-based task matching, cluster-wise knowledge sharing, and continual pre-training of foundation models. These strategies collectively enable scalable, privacy-preserving, and domain-adaptive intelligence across heterogeneous federated environments.
• Figure 5: A taxonomy of techniques for enabling efficient personalized federated intelligence. It is organized into two main categories: (1) Efficient adaptation strategies, including LoRA, prompt tuning, and adapter tuning; and (2) Communication- and computation-efficient methods, such as model pruning and compression.