Source-linked AI summary
A Survey on LLM-powered Agents for Recommender Systems
Qiyao Peng, Hongtao Liu, Hua Huang, Qing Yang, Minglai Shao
TL;DR
Traditional recommender systems have limited ability to understand complex intents, support meaningful interaction, and explain recommendations. This survey systematically reviews LLM-powered agent recommenders by organizing methods into three paradigms, analyzing a four-module architecture, and comparing datasets and evaluation frameworks. It concludes by identifying architectural, evaluation, privacy, and security challenges alongside future research directions.
Problem
Traditional recommendation methods struggle with complex user intents, meaningful interaction, and interpretable explanations, motivating systematic study of LLM-powered agents.
Method
The survey categorizes methods into three paradigms, analyzes Profile, Memory, Planning, and Action modules, and reviews datasets and evaluation methodologies.
Results
The survey provides a systematic account of LLM-powered agent recommendation methods, benchmark datasets, evaluation methodologies, and three future research directions.
Takeaways & Limitations
LLM-powered recommender research can be organized by method objective and agent architecture while requiring broader evaluation frameworks and continued study of system challenges.
Takeaways & Limitations
Current systems still face insufficient integration with traditional recommendation methods, multi-agent collaboration and interpretability challenges, and a lack of unified evaluation standards.
Abstract
from arXiv · showhide
Recommender systems are essential components of many online platforms, yet traditional approaches still struggle with understanding complex user preferences and providing explainable recommendations. The emergence of Large Language Model (LLM)-powered agents offers a promising approach by enabling natural language interactions and interpretable reasoning, potentially transforming research in recommender systems. This survey provides a systematic review of the emerging applications of LLM-powered agents in recommender systems. We identify and analyze three key paradigms in current research: (1) Recommender-oriented approaches, which leverage intelligent agents to enhance the fundamental recommendation mechanisms; (2) Interaction-oriented approaches, which facilitate dynamic user engagement through natural dialogue and interpretable suggestions; and (3) Simulation-oriented approaches, which employ multi-agent frameworks to model complex user-item interactions and system dynamics. Beyond paradigm categorization, we analyze the architectural foundations of LLM-powered recommendation agents, examining their essential components: profile construction, memory management, strategic planning, and action execution. Our investigation extends to a comprehensive analysis of benchmark datasets and evaluation frameworks in this domain. This systematic examination not only illuminates the current state of LLM-powered agent recommender systems but also charts critical challenges and promising research directions in this transformative field.
1 Introduction
The survey frames traditional recommender systems as limited in understanding complex intents, supporting interaction, and explaining recommendations. It organizes LLM-powered agent research through a three-paradigm taxonomy, a four-module architecture, comparative datasets and evaluations, and future directions.
- Traditional recommendation methods struggle with complex user intents, meaningful interaction, and interpretable explanations.
- LLM-powered agents are examined as a way to provide more capable recommendation through reasoning and contextual understanding.
- The survey compares methods, benchmark datasets, evaluation methodologies, and emerging evaluation approaches for LLM-powered recommender systems.
- The survey categorizes existing approaches as recommender-oriented, interaction-oriented, and simulation-oriented.
- It analyzes agents through Profile Construction, Memory Management, Strategic Planning, and Action Execution modules.
2 Background
The background contrasts conventional preference prediction with LLM agents that dynamically decompose tasks and support recommendation through modular, closed-loop processing. The proposed agent-centric formulation synthesizes module outputs into predicted item preferences.
- Traditional Recommendation: Traditional recommendation learns a preference function over users and items from an interaction matrix.
- Traditional Recommendation: Traditional methods struggle with complex intents, meaningful preference-eliciting interaction, and transparent explanations.
- LLM as Agent: An LLM-powered agent dynamically decomposes complex tasks into manageable components rather than relying on static prompt-response processing.
- LLM as Agent: The agent architecture comprises Profile, Memory, Planning, and Action modules for user representation, context retention, strategy formation, and execution.
- LLM Agents for Recommendation: The recommendation framework maps user-specific inputs through functional modules and an integration function to produce a predicted preference distribution over items.
- LLM Agents for Recommendation: These modules operate in a closed loop in which interaction data enriches profiles and memory, informs planning, and supports action and feedback collection.
3 Methods
The survey classifies LLM-powered recommendation methods by objective and analyzes them through a four-module architecture. It illustrates direct recommendation, conversational engagement, simulated interactions, and module-specific mechanisms for profiling, memory, planning, and action.
- Method Objectives: Existing methods are classified as recommender-oriented, interaction-oriented, and simulation-oriented approaches.
- Recommender-oriented: Recommender-oriented methods use planning, reasoning, memory, and tools to generate direct recommendations from user behavior.
- Interaction-oriented: Interaction-oriented methods use natural dialogue and explanations to track preferences and make recommendations conversationally.
- Simulation-oriented: Simulation-oriented methods model users, items, and recommender systems as interacting LLM-powered agents.
- Agent Components: The architecture includes Profile, Memory, Planning, and Action modules that respectively represent users and items, manage experience, formulate strategies, and execute decisions.
- Profile Module: Profile construction captures behavioral patterns and preferences to support dynamic, personalized user representations.
- Memory Module: Memory stores historical interactions and contextual experiences so prior preference signals can inform later recommendations.
- Planning Module: Planning generates multi-step recommendation strategies that balance immediate satisfaction, long-term engagement, and item discovery.
4 Datasets and Evaluations
The survey reviews traditional, conversational, and domain-specific datasets alongside sampling practices and diverse evaluation metrics. Evaluation spans standard recommendation accuracy, language quality, reinforcement-learning outcomes, interaction efficiency, and customized conversational criteria.
- Datasets: Traditional datasets provide user-item interactions, timestamps, and review text for evaluating recommendation models across platforms and scales.
- Datasets: Amazon Review datasets support large-scale evaluation, including Books with 10.3M users and 4.4M items and Video Games with 2.8M users and 137.2K items.
- Datasets: MovieLens ranges from MovieLens-100K with 0.9K users and 1.6K items to MovieLens-20M with 138.5K users and 27.3K items.
- Datasets: Conversational datasets include ReDial, Reddit, and OpenDialKG for evaluating seeker-recommender dialogue and recommendation discussions.
- Evaluation Setup: Because LLM calls consume resources and time, some methods evaluate on sampled user subsets rather than complete datasets.
- Evaluation Metrics: Standard evaluation commonly uses NDCG@K, Recall@K, and HR@K, while some methods measure generated-text quality with ROUGE.
- Evaluation Metrics: Long-term engagement evaluations use trajectory length, average and cumulative rewards, while interaction-focused evaluations use Success Rate and Average Turn.
- Evaluation Metrics: Custom frameworks evaluate conversational agents through criteria such as proactivity, economy, explainability, correctness, consistency, efficiency, and believability.
5 Related Research Fields
LLM-based recommender research builds on both LLM recommendation systems and conversational recommender systems, while addressing limitations in traditional methods and smaller language models.
- Most LLM-based recommender methods focus on rating prediction and sequential recommendation.
- Conversational recommender systems form a related research direction for interactive recommendation.
- Attribute-based conversational methods use rigid dialogue patterns, limiting interaction flexibility.
- Generation-based conversational methods suffer from restricted knowledge and poor generalization in small language models.
6 Future Directions
Future research should improve system architecture, evaluation, and security for LLM-powered recommender systems.
- Optimization of System Architecture: Integration between traditional recommendation methods and LLMs remains insufficient, particularly for multi-agent collaboration and system interpretability.
- Refinement of Evaluation Framework: Unified evaluation standards are needed to measure dialogue quality and recommendation effectiveness more accurately.
- Refinement of Evaluation Framework: Future evaluation should include novel metrics alongside privacy and security considerations.
- Security: LLM-powered recommender systems are vulnerable to adversarial attacks, motivating robust detection and multi-agent defense mechanisms.
7 Conclusion
The survey organizes LLM-powered recommender research into three paradigms and analyzes these approaches through a unified architecture, datasets, and evaluation methods. It concludes by identifying future research directions.
- Existing approaches are categorized into recommender-oriented, interaction-oriented, and simulation-oriented paradigms.
- The survey analyzes these paradigms through a unified four-module architecture.
- The survey reviews current datasets and evaluation methodologies for LLM-powered recommender systems.
- The survey identifies three promising directions for future research.