Toward Federated Fine-Tuning of Large Language Models: Progress, Challenges, and Future Directions

Large language models (LLMs) have been extensively used in downstream tasks, and the successful implementation of such tasks typically involves fine-tuning on domain-specific or user-specific data. As these data are generally not centralized but rather distributed among clients, federated learning (FL) of LLMs, also known as FedLLM, has emerged as a promising paradigm to decentralize LLM adaptation. Nevertheless, it is still challenging to directly apply traditional FL to LLM fine-tuning due to the amplification of model scale, client heterogeneity, communication cost, and privacy risks. This survey offers an overview of FedLLM fine-tuning from methodological and evaluation aspects. This paper first presents the fundamentals of FL, LLM fine-tuning, and their combination. Then this paper categorizes representative works into three groups: efficiency, heterogeneity and personalization, and privacy and security. In addition, this paper summarizes commonly used datasets and evaluation benchmarks to show how existing methods are evaluated under different client partitions and task settings. Finally, according to the given review, this paper identifies future directions such as resource-adaptive collaboration, split federated fine-tuning, continual personalization, and privacy-preserving safety alignment. All things considered, this survey takes FedLLM fine-tuning as its main point of attention, which gives a well-structured reference for researchers and practitioners to get to know about this field and find out potential research paths.