Year: 2025
Type: preprint
Abstract: The rapid advancement of large language models (LLMs) has sparked a paradigm shift in recommender system design, transforming traditional discriminative ranking architectures into unified generative frameworks. Conventional multi-stage cascading architectures, comprising retrieval, ranking, and auction, have achieved remarkable industrial success but remain limited by semantic gaps, stage inconsistencies, and feature fragmentation. In contrast, emerging Generative Recommendation Systems (GRS) formulate recommendation as a sequence generation task, leveraging transformer-based architectures and tokenized item representations to unify retrieval, ranking, and reasoning within a single generative backbone. This survey provides the first comprehensive synthesis of recent industrial progress in generative recommendation. We categorize over twenty state-of-the-art systems along four complementary dimensions including modeling paradigm (encoder-only, decoder-only, encoder–decoder), functional scope (retrieval, ranking, end-to-end frameworks), representation space (semantic ID–based, dense and hybrid representation), and training and alignment objectives (no alignment, alignment via reinforcement learning and preference optimization). We further summarize emerging research on scaling laws for generative recommenders and analyze their implications for efficiency, generalization, and model–data scaling behavior. Our analysis reveals three converging trends: (1) the unification of retrieval and ranking under shared generative architectures; (2) the integration of preference-aligned, reward-driven learning objectives; and (3) the rapid adoption of multimodal and cross-domain foundation models. Finally, we identify open challenges, including latency–scalability trade-offs, robustness under distribution shifts, interpretability of generative reasoning, and multimodal integration. Then propose a forward looking roadmap to guide future research and industrial deployment of next-generation generative recommender systems
Authors: Siqi Liang, Yudi Zhang
Institutions: Purdue University West Lafayette, Iowa State University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7111158992