Abstract: Agentic recommendations cast recommenders as large language model (LLM) agents that can plan, reason, use tools, and interact with users of varying preferences in web applications. However, most existing agentic recommender systems focus on generic single-agent plan-execute workflows or multi-agent task decomposition pipelines. Without recommendation-oriented design, they often underuse the collaborative signals in the user-item interaction history, leading to unsatisfying recommendation results. To address this, we propose the Multi-Agent Collaborative Filtering (MACF) framework for agentic recommendations, drawing an analogy between traditional collaborative filtering algorithms and LLM-based multi-agent collaboration. Specifically, given a target user and query, we instantiate similar users and relevant items as LLM agents with unique profiles. Each agent is able to call retrieval tools, suggest candidate items, and interact with other agents. Different from the static preference aggregation in traditional collaborative filtering, MACF employs a central orchestrator agent to adaptively manage the collaboration between user and item agents via dynamic agent recruitment and personalized collaboration instruction. Experimental results on datasets from three different domains show the advantages of our MACF framework compared to strong agentic recommendation baselines.
Introduction
Recommender systems are evolving from static "black-box" predictors into agentic systems—Large Language Model (LLM) agents that can reason, use tools, and interact with users. However, current agentic recommenders often suffer from two primary limitations. First, single-agent architectures struggle to incorporate diverse preference signals effectively. Second, existing multi-agent systems often use generic roles like "Planner" or "Critic," which fail to capture the nuanced "collaborative signals"—the patterns of shared preferences among similar users—that have powered recommendation engines for decades.
To bridge this gap, we introduce Multi-Agent Collaborative Filtering (MACF). This framework treats the mathematical concept of "neighborhoods" in collaborative filtering as a literal group of expert agents who discuss and refine recommendations under the guidance of a central orchestrator.
The MACF Framework
MACF mimics the logic of traditional Collaborative Filtering (CF) within a multi-agent environment. Instead of abstract roles, it creates agents based on the specific data context of the user and query.
Key Components:
- Dynamic Agent Recruitment: For a given query, the system identifies "similar users" and "relevant items" from historical data. These are instantiated as User Agents and Item Agents, each initialized with a profile summarizing their specific preferences or attributes.
- The Orchestrator Agent: This central "conductor" manages multi-round discussions. It issues Personalized Collaboration Instructions to specific agents (e.g., asking a User Agent to critique a list based on their known preferences).
- Multi-Round Discussion & Sufficiency Test: Agents iteratively suggest, critique, and refine candidates. After each round, the orchestrator performs a "sufficiency test" to check if the current recommendation list meets criteria for quantity, relevance, and ordering. If satisfied, the discussion ends; otherwise, it continues.
- Adaptive Tool Use (ATU): Agents can dynamically call retrieval tools during the discussion to resolve uncertainties or find additional evidence.
Experimental Results
The framework was evaluated on three domains from the Amazon Product Dataset: Clothing, Beauty, and Music. Performance was measured using standard metrics, including Hit Ratio (H@10) and Normalized Discounted Cumulative Gain (N@10).
MACF demonstrated substantial improvements over both traditional and agentic baselines. On the Amazon Music dataset, MACF achieved an H@10 of 0.5450, a significant jump compared to the strongest non-MACF agentic baseline, TAIRA (0.3519). Furthermore, ablation studies confirmed that removing "Personalized Instructions" or "Dynamic Recruitment" led to significant performance drops, proving that the collaborative nature of the system is its primary driver of success.
Efficiency and Insights
A major advantage of MACF is its ability to structure reasoning. Rather than dumping raw user history into a single long prompt, MACF transforms that history into active participants. This prevents the LLM from getting "lost" in context and ensures efficient processing.
As shown in the analysis of discussion rounds, the system is highly efficient. The orchestrator's sufficiency test ensures that most recommendations are finalized within just one or two rounds of discussion, minimizing token usage while maintaining high accuracy.
Conclusion
MACF successfully bridges the gap between traditional recommendation algorithms and modern LLM agents. By treating collaborative filtering as a structured conversation among specialized agents, the framework provides superior performance and reasoning capabilities. Future work will explore scaling these discussions to larger agent pools and leveraging the multi-agent dialogue to provide better explainability for end-users.
Summary updated