Rethinking Recommendation Paradigms: From Pipelines to Agentic Recommender Systems
Source: arXiv:2603.26100 | Posted: 2026-03-27 | Category: Agentic Search / Recommender Engines
Overview
This paper challenges the conventional architecture of large-scale industrial recommender systems, which typically rely on a fixed multi-stage pipeline (recall → ranking → re-ranking) that has evolved from collaborative filtering to deep and large pre-trained models. The authors argue that both multi-stage and "One Model" designs are fundamentally static: models remain black boxes, and system improvement depends on manual hypotheses and engineering effort — an approach that struggles to scale under heterogeneous data and multi-objective business constraints.
The AgenticRS Proposal
The paper proposes an Agentic Recommender System (AgenticRS) that reorganizes key modules as agents. A module is promoted to an agent only when it satisfies three criteria:
1. It forms a functionally closed loop 2. It can be independently evaluated 3. It possesses an evolvable decision space
Self-Evolution Mechanisms
For model agents, two self-evolution paths are outlined:
- Reinforcement learning-style optimization in well-defined action spaces
- LLM-based generation and selection of new architectures and training schemes in open-ended design spaces
- Individual evolution: self-improvement of a single agent
- Compositional evolution: evolution over how multiple agents are selected and connected
- Agentic paradigms make retrieval frequency and strategy themselves learnable, rather than fixed pipeline stages
- High-quality instruction data and click/session logs remain critical; synthetic data requires care against knowledge leakage and distribution shift
- The gap between offline metrics and online satisfaction is widening — LLM-as-judge should be cross-validated with human evaluation
- Latency, cost, interpretability, and safety are hard constraints for industrial deployment, not just academic benchmarks
- Agentic Information Retrieval (arXiv:2410.09713)
- AgentX: Towards Agent-Driven Self-Iteration of Industrial Recommender Systems (arXiv:2606.26859)
- Synergizing RAG and Reasoning: A Systematic Review (arXiv:2504.15909)
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via RL (arXiv:2509.24193)
The framework further distinguishes:
A layered inner/outer reward design couples local optimization with global business objectives.
Key Takeaways
Limitations and Open Problems
The paper leaves open questions around evaluation trustworthiness, latency and cost, hallucination and safety in open-ended agent systems, and cross-lingual/multimodal scaling. Engineering teams should also account for index update frequency, embedding version compatibility, A/B test sensitivity, and failure modes (empty retrieval, wrong tool calls, over-generation) when translating the blueprint into production.
Original Abstract
> Large-scale industrial recommenders typically use a fixed multi-stage pipeline (recall, ranking, re-ranking) and have progressed from collaborative filtering to deep and large pre-trained models. However, both multi-stage and so-called One Model designs remain essentially static: models are black boxes, and system improvement relies on manual hypotheses and engineering, which is hard to scale under heterogeneous data and multi-objective business constraints. We propose an Agentic Recommender System (AgenticRS) that reorganizes key modules as agents. Modules are promoted to agents only when they form a functionally closed loop, can be independently evaluated, and possess an evolvable decision space. For model agents, we outline two self-evolution mechanisms: reinforcement learning style optimization in well-defined action spaces, and large language model based generation and selection of new architectures and training schemes in open-ended design spaces. We further distinguish individual evolution of single agents from compositional evolution over how multiple agents are selected and connected, and use a layered inner and outer reward design to couple local optimization with global objectives. This provides a concise blueprint for turning static pipelines into self-evolving agentic recommender systems.