All Roads Lead to Rome: Unveiling the Trajectory of Recommender Systems Across the LLM Era

TL;DR

This survey traces the trajectory of recommender systems through the LLM era, contrasting traditional list-wise methods with LLM-enhanced and conversational approaches, and culminating in the emergence of autonomous LLM-powered recommendation agents. It argues that two evolution paths—enhancing information content via language-enabled interactions and automating pipeline stages with language foundation models—converge toward agents capable of perception, memory, planning, and tool use. The paper details technical milestones, methodologies, and challenges across conventional list-wise systems, LLM-assisted enhancements, and CRS both before and during the LLM era, before outlining open problems and future directions. The core contribution is a cohesive framing of how LLMs transform personalization technologies, offering a roadmap for building flexible, knowledge-rich, and interactive recommender agents with practical implications for personalization interfaces and ecosystem-level deployment.

Abstract

Recommender systems (RS) are vital for managing information overload and delivering personalized content, responding to users' diverse information needs. The emergence of large language models (LLMs) offers a new horizon for redefining recommender systems with vast general knowledge and reasoning capabilities. Standing across this LLM era, we aim to integrate recommender systems into a broader picture, and pave the way for more comprehensive solutions for future research. Therefore, we first offer a comprehensive overview of the technical progression of recommender systems, particularly focusing on language foundation models and their applications in recommendation. We identify two evolution paths of modern recommender systems -- via list-wise recommendation and conversational recommendation. These two paths finally converge at LLM agents with superior capabilities of long-term memory, reflection, and tool intelligence. Along these two paths, we point out that the information effectiveness of the recommendation is increased, while the user's acquisition cost is decreased. Technical features, research methodologies, and inherent challenges for each milestone along the path are carefully investigated -- from traditional list-wise recommendation to LLM-enhanced recommendation to recommendation with LLM agents. Finally, we highlight several unresolved challenges crucial for the development of future personalization technologies and interfaces and discuss the future prospects.

Figures (6)

• Figure 1: Development trajectory of the modern recommender systems. The $x$-axis denotes key technical milestones in artificial intelligence, while the $y$-axis represents various interaction types, leading to five different recommendation paradigms (marked by different colors). Arrows present the general paradigm shift: recommender systems evolve from the lower left corner to the upper right corner, with a decrease in interactive cost and an increase in effective information.
• Figure 2: Example of a conventional list-wise recommendation system.
• Figure 3: The illustrative dissection of LLM-empowered recommendation. LLM-empowered recommender systems can be divided into two categories: LLMs for feature engineering and LLMs for ranking.
• Figure 4: The illustration of conventional list-wise recommendation and conversational recommender systems (CRS).
• Figure 5: A overview of the LLM-based conversational recommendation system (CRS): (a) The prompting-based LLM-CRS showcases an organized composition of interconnected modules, each serving a distinct purpose. The LLM acts as a controller, interacting with off-the-shelf recommenders to manage information about items, user profiles, dialogues, user behaviors and etc., through prompts, as well as enhancing performance via reasoning and reflection. (b) The fine-tuning-based LLM-CRS can be divided into two structural approaches: one aims to jointly optimize LLM and recommender, while the other formulates CRS as a generative recommendation task, tackled primarily through the LLM. Besides, the LLM is also utilized to generate CRS data, addressing data scarcity issues.