FedSLLM: LLM-Derived Semantic Prototypes for Sample Selection in Federated Recommendation

ACL ARR 2026 January Submission2131 Authors

Keywords: Large Language Model, Sample Selection, Federated Learning, Recommender Systems

Abstract: Recently, large language models (LLMs) have demonstrated strong generative capabilities. These advances create new opportunities for improving federated recommendation (FR), which enables distributed model training while preserving user privacy. However, strict constraints on privacy, fairness, and communication overhead leave a research gap in applying LLMs to FR, particularly in addressing ineffective training caused by biased or unrepresentative samples. To this end, we propose FedSLLM, an FR framework leveraging server-side LLMs to generate semantic prototypes that guide clients in selecting the most informative and representative local samples based on semantic relevance and prediction difficulty. This approach enables effective, lightweight, and privacy-preserving sample selection without deploying LLMs on clients or sharing raw data. Extensive experiments on multiple FR backbones and datasets show that FedSLLM consistently improves recommendation performance, especially under low sampling ratios, while reducing the amount of training data required.

Paper Type: Long

Research Area: Information Extraction and Retrieval

Research Area Keywords: Information Extraction,Machine Learning for NLP

Languages Studied: English

Submission Number: 2131