A Survey on Generative Recommendation: Data, Model, and Tasks

📝 Paper Summary

Generative Recommendation Large Language Models (LLMs) Diffusion Models

This survey establishes a comprehensive 'Data-Model-Task' framework to systematize the paradigm shift from discriminative to generative recommendation, incorporating recent advancements in agents, diffusion models, and large recommendation models.

Core Problem

Existing surveys on generative recommendation (up to 2024) are becoming outdated, often overlooking rapid 2025 advancements in agent-based simulation, diffusion models, and tasks beyond simple supervised fine-tuning.

Why it matters:

Traditional discriminative models (matching users to item IDs) struggle with cold-start, lack semantic understanding, and cannot generate creative content
The field is evolving too fast for previous taxonomies (which focused mainly on LLM embeddings or tuning methods) to capture the full scope of agentic and multimodal approaches
Researchers lack a unified roadmap connecting data synthesis, model architecture changes (e.g., scaling laws), and novel generative applications

Concrete Example: A traditional recommender fails to serve a new user with no history (cold start). A generative approach described in this survey can use an LLM agent to simulate this user's behavior based on a profile, synthesizing 'fake' interaction data to train the model before the user ever clicks an item.

Key Novelty

Unified Data-Model-Task Taxonomy

Data-Level: Categorizes methods using generative models for open-world knowledge augmentation, agent-based behavior simulation, and multi-modal data unification
Model-Level: Distinguishes between LLM-based recommenders, Large Recommendation Models (LRMs) that follow scaling laws, and Diffusion-based approaches for generation
Task-Level: Identifies emerging applications enabled by generation, including conversational interaction, explicit reasoning/explanation, and personalized content creation (AIGC)

Architecture

The hierarchical taxonomy of the survey, dividing the field into Data, Model, and Task levels.

Breakthrough Assessment

8/10

Provides a timely and necessary update to the recommendation taxonomy, specifically integrating 2025-era concepts like agentic simulation and LRMs which were absent in prior surveys.

⚙️ Technical Details

Problem Definition

Setting: Generative Recommendation

Inputs: User interactions, user profiles, item content (text, images), and heterogeneous context

Outputs: Generated items (IDs or content), synthetic data for training, explanations, or conversational responses

Comparison to Prior Work

vs. Wu et al. and Lin et al.: This survey extends the scope to include 2025 research, specifically Agent-based simulation and Diffusion models, which were less prominent in 2024 surveys
vs. Zhao et al.: Introduces the 'Task-level' dimension to cover high-level applications like creative generation and reasoning, rather than just training paradigms
vs. Liu et al. [106]: Broader coverage of 'Large Recommendation Models' (LRMs) and their scaling laws, not just multimodal pretraining

Limitations

The field is moving extremely fast; specific SOTA model names mentioned (e.g., Llama-3) may be outdated quickly
The survey focuses on generative paradigms, potentially underrepresenting hybrid methods that still rely heavily on traditional collaborative filtering
Benchmarks for generative recommendation are still maturing, making cross-paper comparisons difficult

Reproducibility

This is a survey paper reviewing over 200 existing works. It categorizes external papers but does not provide a single code repository for a new model. The authors collected papers from top-tier conferences (ICML, ICLR, NeurIPS, KDD, RecSys) and arXiv.

📚 Prerequisite Knowledge

Prerequisites

Recommender Systems (Collaborative Filtering, Matrix Factorization)
Generative AI (LLMs, Diffusion Models)
Deep Learning (Transformers, GNNs)

Key Terms

Generative Recommendation: A paradigm that reframes recommendation as a generation task (predicting the next token/item or synthesizing content) rather than a discriminative ranking task

LRM: Large Recommendation Model—recommendation architectures designed to scale significantly with data and compute, exhibiting emergent abilities similar to LLMs

Diffusion Models: Generative models that learn to reverse a noise-adding process to generate data; used here for synthesizing user interactions or item images

Agent-based Simulation: Using LLM-powered agents to mimic user behavior (browsing, clicking) to generate synthetic interaction data for training recommenders

SFT: Supervised Fine-Tuning—adapting a pre-trained model to a specific task using labeled data

Cold-start: The difficulty of recommending items to new users or recommending new items that have no interaction history

AIGC: AI-Generated Content—content (text, images, etc.) created by generative models

Discriminative Matching: The traditional recommendation approach of scoring the compatibility between a user vector and an item vector

Chain-of-Thought: Prompting technique enabling LLMs to generate intermediate reasoning steps before the final answer