On the Evolution of Federated Post-Training Large Language Models: A Model Accessibility View

2025-08-22

AI Summary

This work addresses the dual challenges of parameter efficiency and privacy preservation in federated learning (FL) for post-training large language models (LLMs) under black-box access—i.e., when only inference APIs are available. Methodologically, it integrates FL frameworks with parameter-efficient fine-tuning (PEFT) techniques to enable distributed, collaborative optimization of proprietary, API-hosted LLMs without exposing model parameters. The paper makes three key contributions: (1) it presents the first systematic taxonomy for black-box Federated LLMs (FedLLM), organized along two axes—model accessibility (white-box, gray-box, black-box) and parameter efficiency; (2) it establishes a unified classification framework, surveys representative approaches, and identifies core technical challenges; and (3) it clarifies the research trajectory for black-box FedLLM advancement, bridging theoretical privacy guarantees with practical deployment requirements. This work lays foundational groundwork for privacy-respecting, scalable LLM personalization in decentralized settings.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Federated Learning (FL) enables training models across decentralized data silos while preserving client data privacy. Recent research has explored efficient methods for post-training large language models (LLMs) within FL to address computational and communication challenges. While existing approaches often rely on access to LLMs' internal information, which is frequently restricted in real-world scenarios, an inference-only paradigm (black-box FedLLM) has emerged to address these limitations. This paper presents a comprehensive survey on federated tuning for LLMs. We propose a taxonomy categorizing existing studies along two axes: model access-based and parameter efficiency-based optimization. We classify FedLLM approaches into white-box, gray-box, and black-box techniques, highlighting representative methods within each category. We review emerging research treating LLMs as black-box inference APIs and discuss promising directions and open challenges for future research.

Problem

• Surveying federated tuning methods for large language models
• Classifying FedLLM approaches by model accessibility levels
• Addressing limitations of inference-only black-box FedLLM paradigms

Innovation

• Black-box FedLLM inference-only paradigm
• Taxonomy categorizing model access optimization
• Classification into white gray black techniques