Multi-Agent Collaborative Filtering: Orchestrating Users and Items for Agentic Recommendations

Yu Xia, Sungchul Kim, Tong Yu, Ryan A. Rossi, Julian J. McAuley

TL;DR

Multi-Agent Collaborative Filtering (MACF) bridges the gap between traditional recommendation algorithms and modern LLM agents. By instantiating similar users and relevant items as specialized agents that "discuss" a recommendation under a central orchestrator, the framework significantly outperforms both traditional collaborative filtering and existing single-agent LLM systems.

What Problem Does This Paper Address?

Recommender systems are evolving from static "black-box" predictors into agentic systems—Large Language Model (LLM) agents that can reason, use tools, and chat with users. However, current agentic recommenders usually fall into two traps:

  1. Single-agent silos: They use one LLM to plan and execute, making it difficult to incorporate diverse preference signals.
  2. Generic multi-agent roles: They use roles like "Planner" or "Critic," which are too abstract to capture the nuanced "collaborative signals" (the idea that similar users like similar things) that have powered recommendation engines for decades.

The authors identify that while traditional Collaborative Filtering (CF) is great at using neighborhood data, it lacks the reasoning and interactivity of LLMs. Conversely, LLM agents are great at reasoning but often ignore the rich "neighborhood" evidence found in user-item interaction histories.

Methods

MACF mimics the logic of Collaborative Filtering within a multi-agent framework. Instead of generic roles, it creates agents based on the specific data context:

  • Dynamic Agent Recruitment: For a given user and query, the system identifies "similar users" and "relevant items" from the history. It then instantiates these as User Agents and Item Agents, each initialized with a profile summarizing their specific preferences or attributes.
  • The Orchestrator Agent: This is the "conductor" of the system. It manages a multi-round discussion, issuing Personalized Collaboration Instructions to specific agents (e.g., "User Agent X, based on your preference for trendy clothes, critique the current list").
  • Multi-Round Discussion & Sufficiency Test: Agents don't just provide a list; they suggest, critique, and refine candidates. After each round, the orchestrator runs a "sufficiency test" to check if the current recommendation list meets criteria for quantity, relevance, and ordering. If it does, the discussion ends; if not, it continues for another round.
  • Adaptive Tool Use (ATU): Agents can call retrieval tools (like searching for similar items or users) dynamically during the discussion to resolve uncertainties.

Datasets

The framework was evaluated on three diverse domains from the Amazon Product Dataset:

  • Amazon Clothing
  • Amazon Beauty
  • Amazon Music

The authors used a simulator-driven evaluation protocol, measuring performance using standard metrics: Hit Ratio (H@10) and Normalized Discounted Cumulative Gain (N@10).

Key Insights and Ideas

  • CF as a Conversation: The most striking idea is treating the mathematical concept of "neighborhoods" in CF as a literal group of experts (agents) who can argue for or against a specific product.
  • Data-Driven Role Specialization: Instead of hard-coding what an agent does, its "personality" and "expertise" are derived directly from the user-item interaction graph.
  • Structured Reasoning over Raw Context: Most LLM recommenders just dump user history into a prompt. MACF structures that history into active participants, which prevents the LLM from getting "lost" in a long context window.
  • Efficiency through Orchestration: The "sufficiency test" ensures the system doesn't waste LLM tokens. The authors found that most recommendations are finalized within 1–2 rounds of discussion.

Key Results

MACF demonstrated substantial improvements over both traditional and agentic baselines:

  • Superior Performance: On the Amazon Music dataset, MACF achieved an H@10 of 0.5450, a massive jump compared to the strongest non-MACF agentic baseline, TAIRA (0.3519).
  • Better than Traditional CF: It consistently beat standard ItemCF and UserCF, proving that LLM reasoning adds value on top of raw collaborative data.
  • Ablation Success: Removing "Personalized Instructions" or "Dynamic Recruitment" led to significant performance drops, confirming that the "collaborative" part of the multi-agent system is what actually drives the results.

Limitations and Future Work

  • Computational Cost: While the authors optimized the number of rounds, running multiple LLM agents is inherently more expensive and slower than traditional matrix factorization or embedding-based retrieval.
  • Scalability of Agents: The paper uses a subset of similar users/items. Scaling this to hundreds of agents in a single "discussion" would likely lead to noise and high latency.
  • Simulator Bias: Like many recent papers in this field, the evaluation relies on a user simulator. Real-world human-in-the-loop testing would be the next logical step to see if users find the "reasoning" provided by these agents actually helpful.
  • Future Direction: The authors suggest exploring how these agentic discussions could be used to provide better explainability to the end-user, telling them why a group of "similar shoppers" recommended a specific item.

Summary updated

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.
BibTeX
@inproceedings{xia2026multi,
  author = {Yu Xia and Sungchul Kim and Tong Yu and Ryan A. Rossi and Julian J. McAuley},
  title = {{Multi-Agent Collaborative Filtering: Orchestrating Users and Items for Agentic Recommendations}},
  booktitle = {The Web Conference},
  year = {2026},
  doi = {10.1145/3774904.3792931},
}

Loading related content...

Figure