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

Yu Qiao, Huy Q. Le, Avi Deb Raha, Phuong-Nam Tran, Apurba Adhikary, Mengchun Zhang, Loc X. Nguyen, Eui-Nam Huh, Dusit Niyato, Choong Seon Hong

TL;DR

This paper introduces "Artificial Personalized Intelligence" (API) as a new paradigm complementing Artificial General Intelligence (AGI). It proposes "Personalized Federated Intelligence" (PFI) as a key approach to adapt large Foundation Models (FMs) to individual users while preserving privacy and efficiency, leveraging Federated Learning (FL) and advanced adaptation techniques.

What Problem Does This Paper Address?

The paper addresses the growing challenge of adapting massive Foundation Models (FMs), like LLMs, to individual user needs while respecting privacy and computational constraints. FMs offer incredible capabilities but are too large and resource-intensive for direct deployment on edge devices. Centralized fine-tuning raises privacy concerns, and traditional Federated Learning (FL) often struggles to achieve true personalization due to data heterogeneity and the need for a single global model. This gap creates a need for a framework that can deliver personalized intelligence without compromising user data or requiring excessive resources.

Methods

The paper proposes Personalized Federated Intelligence (PFI) as a new paradigm. PFI integrates FL and FMs to enable personalized adaptation at the edge. Key methodological pillars include:

  • Efficient Personalization: Techniques like model pruning, quantization, and parameter-efficient fine-tuning (PEFT) methods (LoRA, prompt tuning, adapter tuning) are used to reduce the communication and computational overhead of adapting large FMs on resource-constrained devices.
  • Trustworthy Adaptation: This involves mitigating bias and ensuring fairness, enhancing interpretability through XAI and reasoning methods, reducing hallucinations, and bolstering robustness against adversarial attacks and privacy breaches.
  • Adaptive Refinement: Retrieval-Augmented Generation (RAG) is highlighted as a crucial mechanism for continuous adaptation. RAG integrates external knowledge to ground FMs and enable context-aware generation, supporting both client-side personalization (e.g., local indexing, retriever adaptation) and server-side optimization (e.g., task matching, knowledge sharing).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • API as a Complement to AGI: The paper argues that while AGI aims for universal intelligence, API focuses on intelligence tailored to individual needs, offering a more practical and user-centric AI future.
  • PFI as the Bridge: Personalized Federated Intelligence (PFI) is presented as the core framework to achieve API by combining the privacy benefits of FL with the generalization power of FMs, enabling edge-based personalization.
  • Modular Adaptation is Key: The success of PFI hinges on lightweight adaptation techniques (PEFT, RAG) that allow FMs to be customized without full retraining, making them suitable for distributed, resource-constrained environments.
  • Trustworthiness is Non-Negotiable: For personalized AI to be adopted, it must be fair, interpretable, robust, and protect user privacy. This requires a multi-faceted approach beyond just efficiency.
  • RAG for Dynamic Personalization: RAG is crucial for keeping personalized models up-to-date by allowing them to access and utilize external, evolving knowledge, thus preventing model drift and improving relevance.

Key Results

The paper is a survey and conceptual framework, so it doesn't present new empirical results. However, it synthesizes existing research to highlight the potential of PFI. For example, it references studies showing that:

  • Federated LoRA and P-Tuning can outperform individual local fine-tuning.
  • PEFT techniques can significantly reduce communication costs and resource demands.
  • RAG-based approaches can improve retrieval quality and downstream generation, especially under domain shifts.

Limitations and Future Work

Limitations: * Privacy-Explainability Trade-offs: Balancing strong privacy guarantees with the need for explainability remains a challenge. * Supervision Scarcity: Mitigating hallucinations in decentralized settings can be difficult due to a lack of readily available supervision. * Efficiency-Robustness Trade-offs: Achieving both high robustness and efficiency simultaneously is an ongoing challenge. * Aggregating PEFT Updates: Challenges remain in effectively aggregating adapter or LoRA-based updates, especially with heterogeneous clients.

Future Work: * Meta-PFI: Integrating autonomous capabilities for dynamic user-agent-world interaction loops. * Quantum-Enabled PFI: Leveraging quantum computing for enhanced optimization, security, and efficiency. * Sustainable PFI: Focusing on energy-aware model optimization, carbon-aware scheduling, and renewable energy integration for environmentally responsible deployment. * Dynamically selecting client-specific strategies in decentralized settings.

Summary updated

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.
BibTeX
@misc{qiao2025towards,
  author = {Yu Qiao and Huy Q. Le and Avi Deb Raha and Phuong-Nam Tran and Apurba Adhikary and Mengchun Zhang and Loc X. Nguyen and Eui-Nam Huh and Dusit Niyato and Choong Seon Hong},
  title = {{Towards Artificial General or Personalized Intelligence? A Survey on Foundation Models for Personalized Federated Intelligence}},
  year = {2025},
  eprint = {2505.06907},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.06907},
}
Figure