-------------------------  METAREVIEW  ------------------------
This paper proposes MACF, a multi-agent LLM recommender that explicitly reintroduces classic collaborative filtering structure by instantiating similar users and relevant items as agents coordinated by an orchestrator. Reviewers find the idea timely and highly relevant to WWW, with clear motivation and consistent gains over strong CF, retrieval, and agentic baselines across three Amazon domains. The writing and figures are generally strong, though a dedicated Related Work section (with more up-to-date 2025 coverage) would improve positioning. Main concerns are practicality and rigor: the approach may be computationally heavy, yet cost/latency/LLM-call statistics and deployment considerations are missing; evaluation is offline/simulator-based and may not reflect online settings. Additional clarity is requested on orchestrator recruitment/pruning decisions and closer comparison to AgentCF and related agent-based CF methods.


----------------------- REVIEW 1 ---------------------

SUBMISSION: 628
TITLE: Multi-Agent Collaborative Filtering: Orchestrating Users and Items for Agentic Recommendations
AUTHORS: Yu Xia, Sungchul Kim, Tong Yu, Ryan Rossi and Julian McAuley

----------- Resource paper -----------
SELECTION: no
----------- Overall evaluation -----------
SCORE: 2 (accept)
----- TEXT:
Strenghts:
1. The paper proposes MACF, a meaningful step toward bringing classic CF structure into agentic recommender systems.
2. The methodology is reasonable and clearly described at a high level.
3. Experiments on three Amazon domains show consistent improvements over a strong set of baselines.

Weakness:
1. There is no formal analysis of complexity, convergence, or guarantees, and the paper does not deeply examine failure modes or robustness.
2. The paper does not analyze computational cost, latency, or real-user experience.


----------------------- REVIEW 2 ---------------------

SUBMISSION: 628
TITLE: Multi-Agent Collaborative Filtering: Orchestrating Users and Items for Agentic Recommendations
AUTHORS: Yu Xia, Sungchul Kim, Tong Yu, Ryan Rossi and Julian McAuley

----------- Resource paper -----------
SELECTION: no
----------- Overall evaluation -----------
SCORE: 1 (weak accept)
----- TEXT:
The paper presents an interesting and well-motivated framework that bridges traditional collaborative filtering with LLM-based multi-agent recommendation. The idea of instantiating similar users and relevant items as agents and coordinating them through a central orchestrator is conceptually sound and aligns well with the intuition behind user-based and item-based CF. The dynamic orchestration and personalized agent instructions provide a structured way to incorporate collaborative signals into agentic recommenders, and the empirical results consistently show improvements over strong baselines.

That said, there are some limitations that prevent this from being a strong accept. The proposed approach is computationally heavy compared to standard CF or hybrid “CF + LLM reranker” pipelines, and the paper does not sufficiently discuss latency, cost, or practical deployment considerations. In addition, the evaluation is conducted on offline Amazon datasets with a simulator-driven protocol; while standard, this does not fully reflect real-world, large-scale, or online recommendation settings.

The paper is clearly written, well structured, and visually well presented. The figures and tables are clear and helpful in understanding the workflow. However, given the rapid evolution of agent-based systems over the last year, it would be reasonable to expect a broader and more up-to-date coverage of recent work in the agentic recommendation and multi-agent LLM literature.

Overall, the paper offers a solid and thoughtful contribution with a compelling idea and promising results, but would benefit from deeper discussion of efficiency, realism of evaluation, and stronger positioning with respect to the most recent advances in agent-based systems.