Survey of LLM-Empowered Agents in Recommendation and Search: Towards Next-Generation Information Retrieval

Authors: Yu Zhang, Shutong Qiao, Jiaqi Zhang, Tzu-Heng Lin, Chen Gao, Yong Li Source: arXiv:2503.05659 (March 2025) Category: Survey · Search Surveys · Agentic Search

Key Points

• Scope. The survey unifies recent work on LLM-empowered agents for recommendation and search, mapping how generative models, retrieval pipelines, and sequential decision-making converge into next-generation information retrieval.
• Evolution timeline.
• 2019–2021: BERT-based reranking and Dense Passage Retrieval (DPR) establish the foundations of neural retrieval.
• 2022–2023: Retrieval-Augmented Generation (RAG) and FreshLLM merge retrieval with generation.
• 2024: Conversational and agentic search, plus Generative Recommender Systems (Gen-RecSys), accelerate.
• 2025–2026: Reinforcement-learning-trained search agents, Deep Research workflows, and GraphRAG emerge as new growth areas.
• Modeling paradigms. The survey contrasts dense retrieval (high recall, low latency, suitable as a first stage), late interaction such as ColBERT (higher precision but larger indexes), generative IR such as DSI and docid generation (simplifies cascading but complicates index updates), and agentic search (sequential decision-making with multi-hop reasoning and self-reflection).
• LLM integration modes. RAG, autonomous agents, and tool-use frameworks provide flexibility and explainability, but introduce latency overhead and error propagation across pipeline stages.
• Optimization objectives. Multi-objective learning-to-rank, RLHF, and online learning align models with business goals, while coping with sparse annotations and shifting distributions.
• Evaluation. The survey compares offline metrics (nDCG@10, MRR, Recall@k, Hit@k), online A/B testing, human preference judgments, and LLM-as-judge approaches, noting persistent gaps between proxy metrics and real user satisfaction.
• Engineering considerations. Authors stress p99 latency budgets, cascaded retrieval with early stopping, async reranking, embedding versioning, PII handling, source whitelisting, and per-query token and GPU cost management as prerequisites for industrial deployment.

Taxonomy

| Dimension | Subcategory | Representative Approaches | Strengths | Limitations | |-----------|-------------|--------------------------|-----------|-------------| | Modeling paradigm | Discriminative vs. generative retrieval | Dual encoders, cross-encoders, DSI, GPT indexing | Mature, scalable | Semantic drift, costly index updates | | LLM integration | RAG / Agent / Tool-use | Retrieval augmentation, search agents, API calls | Flexible, explainable | Latency, error propagation | | Optimization objective | Relevance / diversity / freshness | Multi-objective LTR, RLHF, online learning | Business-aligned | Annotation scarcity | | Evaluation | Offline / Online / Human | nDCG, MRR, LLM-as-judge, A/B testing | Comparable | Deviates from real satisfaction |

Benchmarks and Datasets Referenced

• Retrieval corpora: MS MARCO, BEIR, Natural Questions, plus domain-specific and recommendation datasets.
• Metrics: nDCG@10, MRR, Recall@k, Hit@k, human preference, task success rate, latency, and token cost.
• Baselines: BM25, dense retrievers, cross-encoder rerankers, retrieval-free LLMs, and commercial search APIs.

Main Insights for Practitioners

1. Architecture. Cascaded retrieve–rerank–generate pipelines remain dominant, but agentic paradigms increasingly treat the number and strategy of retrieval calls as learnable decisions. 2. Data. High-quality instruction data is as critical as click and session logs; synthetic data must guard against knowledge leakage and distribution shift. 3. Evaluation. Offline gains no longer guarantee online satisfaction; LLM-as-judge must be cross-validated with human audits. 4. Product. Latency, cost, explainability, and safety are non-negotiable constraints for production systems.

Open Problems and Future Directions

• Lack of unified benchmarks and reproducible, non-proprietary testbeds.
• Bias and reliability of LLM-based evaluation.
• Safety, hallucination, and cost control for autonomous agents operating on the web.
• Process-level supervision for multi-hop reasoning chains.
• Joint retrieval-reasoning training and deeper integration with knowledge graphs and structured databases.
• Multilingual and multimodal consistency.
• Fairness and causal constraints in recommendation pipelines.

Related Entries

• A Comprehensive Survey on Reinforcement Learning-based Agentic Search
• A Survey of Conversational Search (Oct 2024)
• A Survey of Model Architectures in Information Retrieval (Jan 2025)
• A Survey on AI Search with Large Language Models (July 2025)
• A Survey on Knowledge-Oriented Retrieval-Augmented Generation (Mar 2025)
• Cross-Modal Retrieval: A Systematic Review of Methods and Future Directions

Glossary

• IR: Information Retrieval
• RAG: Retrieval-Augmented Generation
• LTR: Learning to Rank
• nDCG: Normalized Discounted Cumulative Gain
• Agentic Search: Modeling search as sequential decision-making with tool use and reflection
• Gen-IR: Generative Information Retrieval

Deployment Checklist

| Check | Question | Recommendation | |-------|----------|----------------| | Data | Do training/index pipelines contain PII? How are versions managed? | Partitioned indexes, redaction, versioned embeddings with rollback | | Latency | What is the p99 budget? How many retrieval calls per query? | Cascaded retrieval with early stopping, hot-query cache, async reranking | | Quality | Do offline gains translate to online CTR/satisfaction? | Interleaved experiments, manual audits, citation verification | | Safety | Does open retrieval risk poisoning or bias? | Source whitelisting, adversarial detection, output filtering | | Cost | Tokens and GPU usage per request? | Routing to smaller models, distillation, hybrid sparse-dense retrieval |