A Survey of Foundation Model-Powered Recommender Systems: From Feature-Based, Generative to Agentic Paradigms

TL;DR

This survey synthesizes how foundation models reshape recommender systems by organizing the literature into three integration paradigms—Feature-Based, Generative, and Agentic. It covers data foundations, representation learning, and downstream tasks, highlighting opportunities and trade-offs across paradigms, from efficient embedding-based methods to interactive agents and fully generative pipelines. Through empirical showcase and analysis of explanations, fairness, multimodal data, and in-context learning, the paper identifies key challenges in deployment, scalability, and safety while outlining open research directions. Overall, the work provides a comprehensive roadmap for advancing FM4RecSys, guiding researchers and practitioners toward practical, robust, and explainable next-generation recommender systems.

Abstract

Recommender systems (RS) have become essential in filtering information and personalizing content for users. RS techniques have traditionally relied on modeling interactions between users and items as well as the features of content using models specific to each task. The emergence of foundation models (FMs), large scale models trained on vast amounts of data such as GPT, LLaMA and CLIP, is reshaping the recommendation paradigm. This survey provides a comprehensive overview of the Foundation Models for Recommender Systems (FM4RecSys), covering their integration in three paradigms: (1) Feature-Based augmentation of representations, (2) Generative recommendation approaches, and (3) Agentic interactive systems. We first review the data foundations of RS, from traditional explicit or implicit feedback to multimodal content sources. We then introduce FMs and their capabilities for representation learning, natural language understanding, and multi-modal reasoning in RS contexts. The core of the survey discusses how FMs enhance RS under different paradigms. Afterward, we examine FM applications in various recommendation tasks. Through an analysis of recent research, we highlight key opportunities that have been realized as well as challenges encountered. Finally, we outline open research directions and technical challenges for next-generation FM4RecSys. This survey not only reviews the state-of-the-art methods but also provides a critical analysis of the trade-offs among the feature-based, the generative, and the agentic paradigms, outlining key open issues and future research directions.

Figures (6)

• Figure 1: Three Paradigms of FM-Powered Recommender Systems
• Figure 2: The taxonomy of FM4RecSys from data characteristics to open problems and opportunities. In contrast to previous surveys, our methodology introduces a unique viewpoint for examining the intersection of FM4RecSys from data characteristics to open problems and opportunities, which are detailed in Section \ref{['sec_difference']}.
• Figure 3: Examples of FM embeddings and tokens for RSs.
• Figure 4: An illustration of the generative paradigm for recommendation. User preference inputs (e.g., the profile description, behavior prompts, and task instructions) are utilized to guide the pre-trained foundation models (FM) for RS. The model can be leveraged in a non-tuning manner by directly utilizing its capabilities or via fine-tuning for specific recommendation tasks, producing various forms of generated recommendations such as item generation, explanation generation, and conversation generation.
• Figure 5: Two types of personalized agents in FM4RecSys: (a) Agent as User Simulator and (b) Agent as Recommender System.