A Survey of Generative Search and Recommendation in the Era of Large Language Models

TL;DR

The paper surveys generative search and generative recommendation in the era of large language models, arguing for a unified four-step framework that treats input formulation, identifiers, training, and inference as the core pipeline. It contrasts generative approaches with traditional ML/DL paradigms, reviews a spectrum of identifier strategies (numeric IDs, titles, N-grams, codebooks, multiviews), and discusses both training and inference mechanisms, including constrained generation with Trie and FM-index. The authors also examine the use of LLMs beyond generation (as feature extractors and rankers) and outline open problems such as memory updates for new content, multimodal extensions, and the role of in-context learning, while envisioning content generation as the next information-seeking paradigm. Overall, the work provides a structured, technical perspective on how generative models can transform retrieval and recommendation, highlighting practical tradeoffs and future research directions for scalable, knowledge-rich information access.

Abstract

With the information explosion on the Web, search and recommendation are foundational infrastructures to satisfying users' information needs. As the two sides of the same coin, both revolve around the same core research problem, matching queries with documents or users with items. In the recent few decades, search and recommendation have experienced synchronous technological paradigm shifts, including machine learning-based and deep learning-based paradigms. Recently, the superintelligent generative large language models have sparked a new paradigm in search and recommendation, i.e., generative search (retrieval) and recommendation, which aims to address the matching problem in a generative manner. In this paper, we provide a comprehensive survey of the emerging paradigm in information systems and summarize the developments in generative search and recommendation from a unified perspective. Rather than simply categorizing existing works, we abstract a unified framework for the generative paradigm and break down the existing works into different stages within this framework to highlight the strengths and weaknesses. And then, we distinguish generative search and recommendation with their unique challenges, identify open problems and future directions, and envision the next information-seeking paradigm.

Figures (5)

• Figure 1: Comparison of the three paradigms for search and recommendation, i.e., machine learning-based, deep learning-based, and generative search and recommendation. "rec." denotes "recommendation".
• Figure 2: Major milestones in the development of generative search.
• Figure 3: Major milestones in the development of generative recommendation regarding item identifier.
• Figure 4: Statistics of six representative datasets in search and recommendation. "Int." denotes "Interactions" and "Rec." denotes "Recommendation". The results are based on the LLaMA2 tokenizer touvron2023llama. The user formulation for generative recommendation includes task descriptions and the user's historical interactions with item titles.
• Figure 5: Illustration of next information-seeking paradigm, i.e., content generation.