A Survey on LLM-powered Agents for Recommender Systems
Source: arXiv:2502.10050 | Authors: Qiyao Peng, Hongtao Liu, Hua Huang, Qing Yang, Minglai Shao | Published: 2025-02-14 | Type: Survey
Overview
Recommender systems are essential components of many online platforms, yet traditional approaches still struggle with understanding complex user preferences and providing explainable recommendations. LLM-powered agents offer a promising path forward via natural-language interactions and interpretable reasoning. This survey provides a systematic review of this emerging intersection.
Three Key Paradigms
1. Recommender-oriented approaches — leverage intelligent agents to enhance the fundamental recommendation mechanisms. 2. Interaction-oriented approaches — facilitate dynamic user engagement through natural dialogue and interpretable suggestions. 3. Simulation-oriented approaches — employ multi-agent frameworks to model complex user-item interactions and system dynamics.
Architectural Foundations
The survey analyzes four essential components of LLM-powered recommendation agents:
- Profile construction
- Memory management
- Strategic planning
- Action execution
- Modeling paradigms: discriminative vs. generative retrieval (dual-tower, cross-encoders, DSI) — mature and scalable, but prone to semantic drift and costly updates.
- LLM integration: RAG / agents / tool-use — flexible and interpretable, but suffers latency and error propagation.
- Optimization targets: relevance, diversity, freshness via multi-objective LTR, RLHF, online learning — business-aligned but annotation-scarce.
- Evaluation: offline (nDCG, MRR, Recall@k), online A/B, and LLM-as-judge — comparable, but diverging from true user satisfaction.
- Lack of unified benchmarks; private data limits reproducibility.
- LLM evaluation bias.
- Safety and cost constraints of agentic systems operating on the open web.
- Future directions: fine-grained process supervision, joint retrieval-reasoning training, enterprise metadata governance, multimodal and cross-lingual consistency, deeper integration with knowledge graphs, and causal/fairness constraints for recommendation.
- A Comprehensive Review on Harnessing Large Language Models to Overcome Recommender System Challenges (arXiv:2507.21117)
- A Comprehensive Survey on Cross-Domain Recommendation (arXiv:2503.14110)
- A Comprehensive Survey on Retrieval Methods in Recommender Systems (arXiv:2407.21022)
- A Review of Modern Recommender Systems Using Generative Models (Gen-RecSys), ACM
- A Survey of Generative Search and Recommendation in the Era of Large Language Models (arXiv:2404.16924)
- Researchers: reproduce core comparisons; check whether statistical significance and compute costs are reported.
- Engineers: extract pluggable modules (encoders, re-rankers, planners) and estimate integration cost with existing stacks.
- Product managers: focus on user-perceivable benefits (latency, answer trustworthiness, multi-turn consistency) rather than offline nDCG alone.
It also reviews benchmark datasets and evaluation frameworks for the domain.
Field Context (from the post's analysis)
Timeline: BERT re-ranking and DPR (2019–2021) established neural retrieval; RAG and FreshLLMS (2022–2023) fused retrieval and generation; conversational/agentic search and Gen-RecSys exploded from 2024; RL-trained search agents, Deep Research, and GraphRAG are growth areas for 2025–2026.
Key Takeaways
1. Architecture: cascaded retrieval + re-ranking + generation remains mainstream, but the agentic paradigm makes retrieval strategy itself a learnable decision process. 2. Data: high-quality instruction data and click/session logs are both critical; synthetic data risks knowledge leakage and distribution shift. 3. Evaluation: the gap between offline metrics and online satisfaction is widening; LLM-as-judge needs cross-validation with human assessment. 4. Deployment: latency, cost, explainability, and safety are hard constraints in production — not just academic benchmarks.