Conversational Recommender Systems Using Generative Models (Gen-CRS): A Literature Review

Abstract

Generative models are profoundly impacting how conversational recommender systems (CRS) are conceptualized, designed, and deployed in practice, enabling mixed-initiative, context-aware, and tool-augmented recommendation workflows that go beyond rigid traditional pipelines. This survey provides a comprehensive review of Generative Conversational Recommender Systems (Gen-CRS) from 2018–2025. We organize the field into three analytical layers: (i) architectural and methodological foundations, covering unified, modular, and agentic system designs, strategies for adapting foundation models, and methods for recommendation and response generation, (ii) knowledge and data foundations, detailing how item-level information (structured catalogs, knowledge graphs, unstructured and multi-modal content), user-level signals (short-term preferences, long-term profiles, and persona representations), and dialogue corpora and logs are integrated into generative workflows, and (iii) evaluation methodologies, structured around what is evaluated (output types), which quality dimensions are measured (e.g., task effectiveness, conversational quality, efficiency, user trust, and ethical concerns), how evaluation is conducted (offline metrics, simulation, user studies, online testing), who evaluates (humans, automated metrics, LLM-based judges), and which stakeholders are considered (consumers, item providers, platforms and society). Our analysis highlights both the capabilities and the risks introduced by generative components, including challenges in grounding, catalog fidelity, controllability, safety, and bias. We argue for responsible and transparent system design, emphasizing validated knowledge integration and evaluation protocols that reflect the full complexity of conversational interactions and system behaviors, and we outline open research challenges that must be addressed to develop reliable and trustworthy Gen-CRS. This survey is also informed by, and partly the result of, tutorial feedback we collected at ACM RecSys 2025 in Prague.